GX10 · local AI server
Lessons on local AI on an NVIDIA GB10-class machine: models, agents, documents, voice, video and security — no cloud.
Aisleriot Solitaire on GX10: Play While the Machine Works
Install Aisleriot on a GB10-class machine and play solitaire while a long job runs: the package, picking a game, SSH forwarding and a quick job check.
UPDATED · 03.10.2026Mahjong Solitaire on GX10: Rules and Strategy for Waiting
Install gnome-mahjongg on a GB10-class machine and play mahjong solitaire while a long job runs: the rules, the free-tile rule, layouts and strategy.
UPDATED · 03.10.2026GNOME Mines: Logic and Probability
Play GNOME Mines on Ubuntu 24.04 for Arm64 as a probability exercise: the rules, four techniques and a small Python example for the exact chance of a mine.
UPDATED · 03.10.2026GNOME Sudoku: a Constraint Puzzle
Solve GNOME Sudoku on Ubuntu 24.04 for Arm64: techniques, the link to constraint problems and a checked Python solver with the MRV heuristic.
VERIFIED · 01.10.2026UPDATED · 01.10.2026n8n on GX10: Automation with a Local AI
Run n8n with Postgres and Ollama in Docker on a GX10 (NVIDIA GB10, ARM64): pinned versions, closed ports, a local model and a first working workflow.
VERIFIED · 01.10.2026UPDATED · 01.10.2026n8n for Condominium Management: 8 Scenarios
Eight n8n workflows for a building manager: fees, payments, requests, assembly, reminders, report, export. A local model and a GDPR note included.
UPDATED · 03.10.2026Games on Arm64: Steam, FEX and Proton
How to run Steam and x86 games on an Arm64 GB10-class machine: Canonical's package, the fex_autoinstall script, Proton and an honest look at FPS.
UPDATED · 03.10.2026StarCraft on ARM64 with Box64 and Wine
An attempt to run StarCraft and Battle.net on a GX10 (NVIDIA GB10, ARM64) via Box64 and Wine: what the docs say, what we have not run, where the risk is.
UPDATED · 03.10.2026Heroes of Might and Magic III on GX10 with VCMI
We run Heroes of Might and Magic III on a GX10 (ARM64) with VCMI, an open-source engine: installation, data from a legal copy, mods and a check.
UPDATED · 03.10.2026Open WebUI and Ollama on GX10: Private Chat
Run Ollama and Open WebUI in Docker on a GX10 (NVIDIA GB10, ARM64): a private ChatGPT-style chat, closed ports, local models and your documents.
UPDATED · 03.10.2026LM Studio on a GX10: GUI, Headless Mode and a Local API Server
Run LM Studio on a GX10: desktop app, the headless llmster and the lms CLI, downloading models and a local OpenAI-compatible server, read against the docs.
UPDATED · 03.10.2026NIM on GX10: an LLM Microservice with an OpenAI API
Run NVIDIA NIM for LLMs on a GB10-class server: NGC key, a Llama 3.1 8B container, an OpenAI-compatible endpoint check and common fixes.
UPDATED · 03.10.2026W&B on GX10: Tracking ML Experiments
Track model training on a GX10 with Weights & Biases: runs, metrics and sweeps — and why NVIDIA's data flywheel blueprint is deprecated.
UPDATED · 03.10.2026Agentic AI Safety: Guardrails and Human Approval
Protect an AI agent from prompt injection: NeMo Guardrails plus human approval before irreversible actions. Checked against the docs as of 03.10.2026.
UPDATED · 03.10.2026NIM API: OpenAI Client for a Local Model
Point a familiar OpenAI client at a local NIM container: check the model, first call, streaming, retries and a closed port on a GB10-class server.
UPDATED · 03.10.2026NVIDIA NGC on GX10: CLI, Containers and NIM
Using the NVIDIA NGC catalogue from a GX10 (GB10, ARM64): NGC CLI, API key, nvcr.io containers and a first NIM for a language model, with a memory sum.
UPDATED · 03.10.2026Speech to Text with Riva and Speech NIM on GX10
Speech recognition on your own machine: what changed for Riva, an ASR container with Speech NIM, streaming and whole-file recognition, personal data.
UPDATED · 03.10.2026CUDA-X on GX10: cuDF, cuML and cuOpt on the GPU
Speed up pandas and scikit-learn with no code changes and run the cuOpt routing server on a GX10 (NVIDIA GB10, ARM64), with honest measuring.
UPDATED · 03.10.2026NVIDIA Cloud Functions: GPU functions and a hybrid with GX10
How to package a container as an NVIDIA Cloud Functions function, call it over HTTP and let a GX10 take the daily load while the peak goes elsewhere.
VERIFIED · 01.10.2026UPDATED · 01.10.2026Whisper and 3CX: Call Recordings in Postgres
3CX call recordings to text with Whisper on a local AI server, via n8n into Postgres: legal basis, notice to callers, retention. GDPR Arts. 6 and 13.
VERIFIED · 01.10.2026UPDATED · 01.10.2026Damage Photo to PDF Report with a Local AI Model
A local vision model (Ollama) reads a damage photo and drafts a PDF report: metadata and GPS stripped, a person approves, on a GB10-class server.
VERIFIED · 01.10.2026UPDATED · 01.10.2026Classifying Signals with a Local Model
A local model sorts incoming messages by type, urgency and owner using an Ollama schema; a person reviews the uncertain ones. Fictional data.
UPDATED · 03.10.2026vLLM on GX10: a Language Model Server with an OpenAI-Compatible API
Run vLLM on a GX10 (NVIDIA GB10, ARM64): install with uv or Docker, an OpenAI-compatible server on loopback only, requests, streaming and metrics.
VERIFIED · 01.10.2026UPDATED · 01.10.2026Multi-Agent Request Pipeline with a Human
Five stages on PostgreSQL and n8n: intake, acknowledgement, routing, SLA and a closure that a human always approves. A fictional example.
VERIFIED · 01.10.2026UPDATED · 01.10.2026Fee Reminders: Polite, Scheduled, With a Human
Automatic fee reminders D+1 to D+30: polite tone, workdays, EUR amounts, a legal basis for e-mail/SMS and human approval. Run on fictional data.
VERIFIED · 01.10.2026UPDATED · 01.10.2026S3 Document Storage: Versions and Retention
Your own S3 storage on a local AI server (GB10, ARM64): MinIO from source, versions, Object Lock, alternatives. Checked against sources on 01.10.2026.
VERIFIED · 01.10.2026UPDATED · 01.10.2026Hybrid search: bge-m3 and pgvector in Postgres
Hybrid search by meaning and by words: bge-m3 via Ollama, pgvector in Postgres 18, HNSW and RRF. Fictional documents, no measured numbers.
VERIFIED · 01.10.2026UPDATED · 01.10.2026FastAPI Gateway for Ollama: Keys and Limits
A FastAPI gateway in front of Ollama on a GX10: an API key header, request limits, token streaming through /api/chat and TLS via a reverse proxy.
VERIFIED · 01.10.2026UPDATED · 01.10.2026Choosing a model: memory, speed and quality
How to choose a model for a local AI server: memory, the speed ceiling set by memory bandwidth, and quality, with a calculator and three invented cases.
VERIFIED · 01.10.2026UPDATED · 01.10.2026Appsmith Dashboard over Postgres and Ollama
An Appsmith dashboard over Postgres and local Ollama in Docker on a GX10 (NVIDIA GB10, ARM64): pinned versions, closed ports, a morning briefing.
UPDATED · 03.10.2026RTSP Video Intake from IP Cameras on GX10
Connect IP cameras to a GX10 over RTSP: an FFmpeg test, a Python class with reconnect, keyframes on disk and records in PostgreSQL.
UPDATED · 03.10.2026Object Detection with YOLO11 on GX10
Detect people and vehicles in real time with YOLO11 and TensorRT on a GX10: guarded zones, several RTSP cameras and logging to PostgreSQL.
UPDATED · 03.10.2026Zone Alerts from Video Cameras on GX10
Turn events detected by cameras into timely alerts: guarded zones, a cooldown between signals with Redis, n8n and incident logging in PostgreSQL.
UPDATED · 03.10.2026llama.cpp on GX10: a Light AI Server
Build llama.cpp with CUDA on a GX10 (NVIDIA GB10, ARM64), run llama-server with an OpenAI-compatible API and call it from Python. Checked as of 03.10.2026.
UPDATED · 03.10.2026Video Archive Search with AI on GX10
Find a moment in video recordings with a description in plain language: image embeddings with OpenCLIP, pgvector in PostgreSQL and a small FastAPI service.
UPDATED · 03.10.2026Incident Report from a Spoken Description on GX10
A spoken incident description becomes a draft report on a GX10: Whisper, structuring with Ollama, a PDF from HTML, review and signature by a person.
UPDATED · 03.10.2026Event Dashboard: Grafana and TimescaleDB on GX10
A security event dashboard on a GX10: TimescaleDB, Grafana, SQL panels and an alert to n8n. No frames or faces, with a GDPR legal note.
UPDATED · 03.10.2026DeepStream: Multiple Cameras on GX10
Process several video streams at once with NVIDIA DeepStream 9.1 in Docker on a GX10: batching, a test run, messages to a broker and a legal note.
UPDATED · 03.10.2026License Plate Recognition on GX10: Detector and OCR
Detect and read license plates on a local GX10 AI server: a detector, OCR, data minimisation, a short retention period and the legal framework.
UPDATED · 03.10.2026Camera Anomalies on GX10 Without Profiling
Detect an unusual number of objects in a zone on a local GX10 AI server with PostgreSQL statistics, without storing frames or profiling people.
UPDATED · 03.10.2026Face Access Control and GDPR on GX10
How to build face access control on a local GX10 AI server with consent, retention limits and an event log, and what GDPR and the model licence say.
VERIFIED · 01.10.2026UPDATED · 01.10.2026Invoice OCR on GX10: Nemotron, EIK and VIES
Run Nemotron OCR v2 on a GB10-class server, read invoices, check the Bulgarian EIK checksum and VAT numbers in VIES. Checks tested; Bulgarian OCR not.
VERIFIED · 01.10.2026UPDATED · 01.10.2026Chart of Accounts Expenses: LoRA Fine-Tuning on GX10
Fine-tune a Bulgarian model with LLaMA-Factory and LoRA on a GX10 (NVIDIA GB10) to classify expenses by chart of accounts class 6. Fictional data, evaluation.
VERIFIED · 01.10.2026UPDATED · 01.10.2026VAT Journals and Return: a Draft for the Accountant
Draft Bulgarian VAT journals and return: Python checks and totals, the tax authority file format and the 14th-day deadline. An accountant files.
UPDATED · 03.10.2026Nemotron Nano on GX10: a Local Model
Run Nemotron 3 Nano with Ollama on a GX10 (NVIDIA GB10): pick 4B or 30B, switch thinking on and off and call it through the API. Checked 03.10.2026.
VERIFIED · 01.10.2026UPDATED · 01.10.2026A Local Accounting Assistant That Cites and Refuses
A local assistant over the Bulgarian CIT and VAT Acts: it searches the official text, quotes the article with its edition and refuses when nothing fits.
UPDATED · 03.10.2026Invoices from PDF to Draft Entry: a Local Pipeline
A local invoice pipeline on GX10: PDF to image, OCR, model-extracted fields, checks on totals and VAT, and a draft entry for human approval.
VERIFIED · 01.10.2026UPDATED · 01.10.2026Bank Reconciliation: Match Movements to Invoices
Match bank movements (MT940, CAMT.053) to invoices in Python: EUR with Decimal, fuzzy names with rapidfuzz, human approval. Run on fictional data.
VERIFIED · 01.10.2026UPDATED · 01.10.2026Corporate Tax Scenarios with a Local Model: a Draft for the Adviser
A local language model drafts corporate income tax scenarios for a fictional company in euro; code does the maths and an accountant or tax adviser decides.
VERIFIED · 01.10.2026UPDATED · 01.10.2026Annual Accounts from Postgres to Excel: a Draft for the Accountant
Draft annual financial statements: a trial balance from PostgreSQL into Excel with Python and openpyxl, a balance check and review by an accountant.
VERIFIED · 01.10.2026UPDATED · 01.10.2026Contract Review with OCR and a Local AI
Turn a scanned or text contract into a preliminary HTML report: OCR with Tesseract, a local model via Ollama, quotes to verify, GDPR notes and a lawyer's review.
VERIFIED · 01.10.2026UPDATED · 01.10.2026Searching contracts with citations: a local RAG
A local RAG over contracts with Qdrant, bge-m3 and Ollama: answers with a verified quote and a refusal when not found. Fictional contracts and GDPR.
VERIFIED · 01.10.2026UPDATED · 01.10.2026Notary Deed OCR: Property Data Without the ID Number
A notary deed scan becomes a property record: Tesseract OCR, a local model, the ID number masked before the model and never stored. GDPR Art. 5, 87.
VERIFIED · 01.10.2026UPDATED · 01.10.2026Contract deadlines: from a clause to a reminder
From a contract clause to a date and a reminder: Art. 72 ZZD periods, 2026–2027 holidays, human confirmation. Not legal advice.
UPDATED · 03.10.2026Legal Assistant with Citations: Local RAG
Local RAG over official law texts: pgvector, hybrid search, answers with citations and a refusal when nothing is found. Not a substitute for a lawyer.
UPDATED · 03.10.2026TensorRT-LLM on GX10: a Language Model Server from an NVIDIA Container
Run TensorRT-LLM on a GX10 (NVIDIA GB10, ARM64) from an NVIDIA container: a check, an example model and an OpenAI-compatible server on loopback only.
VERIFIED · 01.10.2026UPDATED · 01.10.2026Draft Contracts from Templates and Clauses
Jinja2 templates and a RAG library of clauses under Bulgarian contract law make a Word draft. Draft only: a lawyer reviews and approves. Not legal advice.
UPDATED · 03.10.2026Civil Procedure Deadlines: from the Notice to the Reminder
Calculate Bulgarian civil procedural deadlines (GPK) with plain tested code on a local AI server; a lawyer confirms, reminders follow. Not legal advice.
VERIFIED · 01.10.2026UPDATED · 01.10.2026GDPR Assistant: Searching the Regulation, a Processing Agreement and a Breach Tracker
A GDPR preparation aid: search the official text with verbatim quotes, draft a processing agreement, track the 72-hour breach window. Not legal advice.
UPDATED · 03.10.2026Notary Deeds: A Register and Data Checks on GX10
A minimal register of property data from notary deeds in PostgreSQL: validation, checks for duplicates and incomplete records, review by a person.
UPDATED · 03.10.2026Ajax Alarm Events and a Local AI: Event Triage
Triage events from an alarm system such as Ajax with a local AI: rules first, a model with schema output, a human in the loop and a way to measure errors.
UPDATED · 03.10.2026A Live Multi-Site Dashboard: WebSocket and PostgreSQL NOTIFY
A dashboard for many sites: alerts and status from each site, WebSocket to the browser, PostgreSQL NOTIFY, camera frames on request and reports in Grafana.
UPDATED · 03.10.2026CCTV DPIA: A Draft with a Local AI
How a local AI on a GX10 helps prepare a draft data protection impact assessment (DPIA) for CCTV under Article 35 of the GDPR, for a lawyer to review.
UPDATED · 03.10.2026Patrol Routes with OR-Tools and PostGIS
Plan security patrol routes on a local AI server: zones in PostGIS, routes with OR-Tools and an LLM assistant that proposes while a human decides.
UPDATED · 03.10.2026Shift Scheduling with CP-SAT and PostgreSQL
Build a weekly shift roster with OR-Tools CP-SAT and PostgreSQL: rules as data, checks in the database, an LLM assistant for explanations and substitutes.
UPDATED · 03.10.2026An Incident Classifier with Redis Streams and a Local Multimodal Model
An incident classifier: Redis Streams with consumer groups, a local multimodal model with schema output, an audit table and a human who decides.
UPDATED · 03.10.2026SGLang on GX10: a Server with Schema-Constrained Output and a Prefix Cache
Run SGLang on a GX10 (NVIDIA GB10, ARM64) in a container: an OpenAI-compatible server on a closed port, JSON schema, a prefix cache and an n8n example.
UPDATED · 03.10.2026A Monthly Report with a Local AI: From Statistics to Word and Excel
Automate a monthly report: PostgreSQL computes the numbers, a local model via Ollama drafts the text, Word and Excel fill themselves. A person approves.
UPDATED · 03.10.2026Camera Health Monitoring and Predictive Maintenance
A general model for monitoring camera health: an RTSP probe, frame sharpness, TimescaleDB, deviation from the norm and a draft ticket from a local model.
UPDATED · 03.10.2026Smart Building: Sensors, TimescaleDB and Dashboards on GX10
A smart building on GX10: sensors over MQTT, TimescaleDB, Grafana dashboards and alarms. An invented public building, with a GDPR note on presence.
UPDATED · 03.10.2026Event Planning Assistant with a Local AI on GX10
An event-planning assistant on GX10: halls in ChromaDB, availability through CalDAV, four roles and a WebSocket chat. All data in the lesson is invented.
UPDATED · 03.10.2026Drafting an EU Grant Proposal with a Local AI on GX10
How a local AI on GX10 helps draft an application for an EU call: official documents, sections by the template, budget arithmetic, consultant review.
UPDATED · 03.10.2026Visitor Analytics for a Cultural and Sports Complex: Counting Entrances Without Tracking People
Count the entrances of a cultural and sports complex without tracking people: DeepStream, TimescaleDB, a forecast, Grafana and a GDPR box.
UPDATED · 03.10.2026Venue RAG Assistant on a Local GX10
A venue assistant on GX10: hall knowledge in ChromaDB with bge-m3, availability via an Ollama function call, a closed chat. All data is invented.
UPDATED · 03.10.2026Ticket Demand Forecast with Prophet and a Local AI
A ticket-sales forecast with Prophet on synthetic data, an evaluation against a naive forecast and price ideas from a local AI. No buyer data is used.
UPDATED · 03.10.2026Audience Segmentation Without Profiling People
Groups of ticket-purchase patterns with k-means on aggregated synthetic data and drafts of general messages from a local AI. No profiling of people.
UPDATED · 03.10.2026Speculative Decoding on GX10: Faster Model Output
How a draft model speeds up a language model on GX10: SGLang and vLLM per current docs, a memory calculation and an honest before-and-after measurement.
UPDATED · 03.10.2026Marketing Content with a Local AI on GX10
Marketing helper for a cultural and sports complex on GX10: copy with Llama 3.3 70B, a FLUX.1 schnell poster and human approval. All data is invented.
UPDATED · 03.10.2026Access Auditing: GDPR and Employment Law
How to log sensitive actions in systems on GX10 without profiling people: GDPR, employment law, pseudonyms, human review, a lawyer's advice.
UPDATED · 03.10.2026GDPR Art. 28 Contracts with a Local AI
A Postgres register of data processors, draft GDPR Art. 28 contracts with a local model and n8n reminders. Full EUR-Lex text. Not a lawyer substitute.
UPDATED · 03.10.2026Smart Building: Sensors and Climate on GX10
Temperature and CO₂ sensors over MQTT into TimescaleDB, climate suggestions and anomaly explanations with a local model on a GX10. All data is invented.
UPDATED · 03.10.2026Ticketing Chatbot with a Local AI on GX10
A ticketing chatbot for a cultural and sports complex: FastAPI and WebSocket, intent detection with a local model, Postgres availability, human handoff.
UPDATED · 03.10.2026Cultural Programme Curator: Gaps, Balance and Similar Artists with a Local AI
Find free dates, check the genre balance and suggest similar artists with pgvector and a local model through Ollama. A person decides in the end.
UPDATED · 03.10.2026Revenue Analysis with a Local AI: Deviations, Year on Year and a Summary for the Manager
Monthly revenue by stream with SQL, a comparison with budget and last year, reasons from people and a draft summary from a local model. A person approves.
UPDATED · 03.10.2026HVAC with Machine Learning: Forecast, Suggestion, Approval
Forecasting the HVAC load of a hall with scikit-learn on GX10: a suggestion, safety limits and human approval. The data in the lesson is simulated.
UPDATED · 03.10.2026Knowledge Atoms: Automatically Splitting Text into Self-Contained Units
How to split transcripts and documents into self-contained knowledge atoms: Whisper, a local model, bge-m3, Chroma, and duplicate and quality checks.
UPDATED · 03.10.2026NVFP4 on GX10: Quantizing Language Models
What NVFP4 is, how much memory it saves, how to quantize a model with Model Optimizer and serve it with vLLM on an NVIDIA GB10-class machine (Blackwell).
TESTED · 3 Oct 2026UPDATED · 03.10.2026A CRM in PostgreSQL with a Local AI: Deal Stages and Email Drafts
A small CRM in PostgreSQL: deal stages with history, a local model drafts an email, a person approves and sends it. Includes a GDPR box for contact data.
TESTED · 3 Oct 2026UPDATED · 03.10.2026A Proposal from a Template with a Local AI: the Code Calculates, the Model Writes the Text
A Word proposal from a template: code calculates sums and VAT, a local model writes only text, a person approves. Not legal or financial advice.
VERIFIED · 01.10.2026UPDATED · 01.10.2026Bulgarian SAF-T from PostgreSQL: XML, XSD Validation and n8n
Prepare a Bulgarian SAF-T file from PostgreSQL with Python and lxml, validate it against an XSD and use n8n for the deadline reminder. Law quoted.
UPDATED · 03.10.2026Questions to an Office's Documents: a Local RAG with Citations
A local assistant that answers questions about an office's documents with citations and a refusal when nothing is found: bge-m3, pgvector, a local model.
TESTED · 3 Oct 2026UPDATED · 03.10.2026Monthly Accounting Report with a Local AI: from Journal to PDF
SQL turns an accounting journal into a PDF pack, a local model drafts the commentary, a checksum guards the file and a person approves it.
UPDATED · 03.10.2026Deadlines from Documents into an .ics Calendar with a Local AI
A local model reads contracts and invoices, finds due dates and saves them to an .ics calendar with reminders. Python, Ollama, FastAPI on a GX10.
UPDATED · 03.10.2026VLMs on GX10: Models That See and Read
Run a local VLM on GX10: Qwen2.5-VL sizes in Ollama, sending an image from the command line and from Python, documents and charts, human review rules.
TESTED · 3 Oct 2026UPDATED · 03.10.2026Preliminary Contract vs Notarial Deed: Comparing Fields on GX10
A model extracts fields with quotes from a preliminary contract and a notarial deed, and code compares them: price, area, encumbrances. For a lawyer.
UPDATED · 03.10.2026Legal RAG with Qdrant: Laws and Court Rulings
Local legal RAG with Qdrant and Ollama: official laws and court rulings, cited answers, citation check, refusal if nothing found. Not legal advice.
TESTED · 3 Oct 2026UPDATED · 03.10.2026From a Record to a Draft PDF Form: Filling with pypdf on GX10
Filling a draft of a PDF form from a minimal property record with pypdf: refusal on missing fields, read-back check, no personal data and no submission.
TESTED · 3 Oct 2026UPDATED · 03.10.2026A Case Brief from Many Documents: Map-Reduce with a Local AI on GX10
Map-reduce over files with a local model: a summary of each document with quotes, a filter for untrue quotes and a draft for a lawyer or consultant.
UPDATED · 03.10.2026A Monthly Operations Report for Sites with a Local AI on GX10
A monthly report from incidents, patrols, alarms and cameras on GX10: SQL indicators, an Ollama comment, a draft PDF and human approval.
UPDATED · 03.10.2026NIS2 and Suppliers: Draft Answers from Verified Facts Only
How a supplier drafts answers to a NIS2 security questionnaire from verified facts only, with automatic checks and human review. Not legal advice.
TESTED · 3 Oct 2026UPDATED · 03.10.2026A Financial Model with Scenarios on GX10: pandas, Forecast and cuDF Acceleration
A small financial model with scenarios on GX10: pandas and scikit-learn, a checked forecast and speed-up with cudf.pandas. Not financial advice.
TESTED · 3 Oct 2026UPDATED · 03.10.2026Patterns in Incidents: Grouping Aggregated Data
Clustering incidents by hour, weekday and site on GX10: KMeans, DBSCAN, a control run and an honest reading of the result. No profiling of people.
TESTED · 3 Oct 2026UPDATED · 03.10.2026Budget vs Actual: a Report with SQL and a Local AI
A budget-vs-actual report for an organisation with a subsidy: SQL calculates the variances, a local model drafts a comment, a person approves.
UPDATED · 03.10.2026LLaMA Factory on GX10: Fine-Tuning with LoRA
Train a model with LoRA on a GX10 (NVIDIA GB10, ARM64) using LLaMA Factory: data, YAML setup, the web UI, merging and export to Ollama.
UPDATED · 03.10.2026A Public Procurement Check with a Local AI
A preparation helper: thresholds from your own file, article search in the law, a local model with citations. Not legal advice. Ollama, pgvector.
UPDATED · 03.10.2026Parking with AI: Sessions and Occupancy with Minimal Personal Data on GX10
A lot of about 200 spaces on GX10: entry and exit, occupancy, retention and plates as personal data. Nobody is charged; everything is invented.
UPDATED · 03.10.2026Sanctions Lists: Screening with a Local AI
How to check counterparties against the EU, UN and OFAC sanctions lists: fuzzy matching, a local model for a second look and a human who decides.
VERIFIED · 01.10.2026UPDATED · 01.10.2026A ZChOD and Labour Code Assistant with Guardrails
A local assistant for the Bulgarian private-security act and the Labour Code: RAG with NeMo Guardrails, cited articles, invented-article check, disclaimer.
UPDATED · 03.10.2026Keys and Access Cards: Request, Risk Score and Human Approval
A key and card log on a local server: risk score, human approval through n8n, reminders and a protected audit log. Made-up data only.
UPDATED · 03.10.2026Checking Shift Schedules: SQL Views and a Local Model
Check a work schedule before publishing: SQL views for shifts, rest, weekly and night hours, limits in a table, a local model and n8n.
UPDATED · 03.10.2026EU Grant Eligibility Checks with Local AI
A local RAG on GB10: official EU programme guides in pgvector, a verdict with a quote from the guide, deadlines by source and clear limits of the method.
UPDATED · 03.10.2026Conference Revenue Forecast and a PDF Report
An honest conference revenue forecast as a range: PostgreSQL, Python and a PDF report with WeasyPrint. Made-up numbers. Not financial advice.
UPDATED · 03.10.2026Tenant Health Score: an Aggregate View
A 0–100 health score for groups of leases by month, built on invented data with PostgreSQL and Grafana: no profiling of people, no automated decisions.
UPDATED · 03.10.2026Unsloth on GX10: Faster Fine-Tuning with LoRA
Train a LoRA adapter with Unsloth on a GX10 (NVIDIA GB10, ARM64): the DGX Spark container, data, a Python script, export to GGUF and Ollama.
UPDATED · 03.10.2026Stakeholder Briefing AI: Briefings by Audience
A local model via Ollama assembles a structured briefing for an audience role and saves it as a Word draft for human review. No real people involved.
UPDATED · 03.10.2026Danube and Black Sea: a Coordinator for EU Programmes with Local RAG
A local assistant answers questions on EU strategies and programmes for the Danube and Black Sea regions from official documents only, with citations.
UPDATED · 03.10.2026Audit Log on GX10: A Narrative with Local AI
Build an audit log on a GB10-class local AI server: SQL finds the anomalies, a local model describes them, and the PDF carries a SHA-256 fingerprint.
UPDATED · 03.10.2026NemoClaw: Local Model and Telegram
After installing NemoClaw: how to pick a local model for 128 GB of shared memory, connect Telegram for yourself only, and check the limits of the sandbox.
UPDATED · 03.10.2026Three GB10 Machines in a Ring: a QSFP Cluster
Connect three GB10 machines in a ring with QSFP cables: network interfaces, SSH keys and an NCCL test, following the NVIDIA guide as of 03.10.2026.
UPDATED · 03.10.2026cuTile and TileGym: GPU Kernels in Python on GX10
Run NVIDIA TileGym benchmarks on a GX10 (GB10): cuTile kernels, swapping kernels in a language model and Flash Attention. Checked 3 Oct 2026.
VERIFIED · 01.10.2026UPDATED · 01.10.2026PaddleOCR-VL on GX10: documents to Markdown
PaddleOCR-VL on a local NVIDIA GB10-class server: how it works, how to serve it with vLLM, what applies to Arm and Blackwell, and how to read benchmarks.
VERIFIED · 01.10.2026UPDATED · 01.10.2026Qianfan-OCR on GX10: Documents to Markdown
Qianfan-OCR from Baidu: one model turns a page into Markdown or JSON. Checked against the official page: licence, size, languages, Layout-as-Thought.
VERIFIED · 01.10.2026UPDATED · 01.10.2026GLM-OCR on GX10: an MIT-licensed OCR model
GLM-OCR (0.9B, MIT) on a local NVIDIA GB10-class server: licence, languages, transformers and vLLM, sourced benchmarks and what we have not run.
VERIFIED · 01.10.2026UPDATED · 01.10.2026DeepSeek-OCR-2 on GX10: long documents
DeepSeek-OCR-2 (3B, Apache-2.0) on a local NVIDIA GB10-class server: how it works, how to run it with vLLM and transformers, and what is still unverified.
UPDATED · 03.10.2026NeMo on GX10: Fine-Tuning Models with AutoModel
What NeMo is today and how to tune a language model with NeMo AutoModel (LoRA, QLoRA, full) on a GX10 (NVIDIA GB10): container, commands, memory limits.
VERIFIED · 01.10.2026UPDATED · 01.10.2026MinerU on GX10: documents to Markdown
MinerU 4.0 on a local NVIDIA GB10-class server: PDF and Office files to Markdown and JSON for language models, tiers, licence and unchecked points.
VERIFIED · 01.10.2026UPDATED · 01.10.2026dots.ocr on GX10: Page Parsing and SVG
dots.ocr and its successor dots.mocr on a local NVIDIA GB10-class server: page parsing with vLLM, tables, diagram-to-SVG and what has not been verified.
VERIFIED · 01.10.2026UPDATED · 01.10.2026HunyuanOCR on GX10: a licence that excludes the EU
HunyuanOCR 1.5 (1B, Tencent) reads documents with coordinates, but its licence excludes the EU. What we checked, what it is, and what to use instead.
VERIFIED · 01.10.2026UPDATED · 01.10.2026Tesseract: Cyrillic OCR Without a GPU
Tesseract as a cheap CPU fallback OCR: Bulgarian text with bul, --psm modes, choosing data (fast, best), pytesseract and a threshold for escalating to a heavier model.
VERIFIED · 01.10.2026UPDATED · 01.10.2026EasyOCR on GX10: Simple Cyrillic OCR
Simple Cyrillic OCR with EasyOCR on an NVIDIA GB10-class server: PyTorch for ARM64 with CUDA, the bg language code, a small example and honest limits.
VERIFIED · 01.10.2026UPDATED · 01.10.2026PaddleOCR on GX10: PP-OCR and PP-Structure
Classic PaddleOCR 3.x on a local NVIDIA GB10-class server: PP-OCR for text, PP-StructureV3 for tables, Cyrillic, and what changed since version 2.x.
VERIFIED · 01.10.2026UPDATED · 01.10.2026OCR 2026 on GX10: which tool for what
Which-for-what matrix for 11 OCR tools on a local NVIDIA GB10-class server: Bulgarian, licences, CPU or GPU. Checked against official sources, 01.10.2026.
UPDATED · 03.10.2026Nemotron 3.5 Lightning and Switchyard: a Cheap Agent, Smart Routing
Run Nemotron 3.5 Lightning locally on a GB10-class machine and learn how NeMo Switchyard routes requests between a local and a cloud model.
UPDATED · 03.10.2026RTX Spark and CUDA on Windows on Arm
What RTX Spark is, what the CUDA 13.4 Developer Preview brings to Windows on Arm, how it differs from the GB10 in GX10, and when each platform makes sense.
VERIFIED · 01.10.2026UPDATED · 01.10.2026What fits on GX10: the arithmetic before you download
How much memory does a model need? Weights, KV cache and overhead on a GX10, recomputed from real model configs, plus the MoE trap and a working ceiling.
UPDATED · 03.10.2026PyTorch on GB10: BF16, torch.compile and Profiling
PyTorch on a GB10-class local AI server: an NVIDIA container, BF16 mixed precision, torch.compile and profiling, measured rather than promised.
VERIFIED · 01.10.2026UPDATED · 01.10.2026Which model is which version — a reference checked on 01.10.2026
Which model is which version: Claude, Qwen, Gemma, DeepSeek, GLM, Kimi, FLUX and OCR models checked on 01.10.2026 with licences; unverified rows flagged.
UPDATED · 03.10.2026FLUX.1 LoRA on GX10: Your Own Style and Licence
Train a LoRA on FLUX.1 [dev] with SimpleTuner on a GX10 (NVIDIA GB10) and check the licence: such a LoRA is a derivative, not for commercial work.
UPDATED · 03.10.2026ComfyUI on GX10: Image Generation with Nodes, and Which Models Are Fine for Paid Work
Run ComfyUI on a GX10 (NVIDIA GB10, ARM64) with FLUX and Z-Image-Turbo, and check the licences of image models: which ones are fine for paid work.
UPDATED · 03.10.2026CLI Coding Agent on GX10: Aider and Ollama
A local coding helper in the terminal: Aider and Claude Code with Ollama on a GX10 (NVIDIA GB10): setup, safe habits and a small agent of your own.
UPDATED · 03.10.2026Vibe Coding in VS Code with a Local AI
Set up VS Code with Continue and Ollama on a GX10: autocomplete, chat, edit and agent from a local model, no cloud. Checked against the docs on 03.10.2026.
UPDATED · 03.10.2026Hermes: A Reflective Agent on a Local AI
A small agent with plan, action, reflection and memory on Ollama on a GX10: restricted tools, SQLite memory, and new tools only after human review.
UPDATED · 03.10.2026Multi-Agent Chatbot with a Router on GX10
A chatbot with a router and specialists on a local AI server: a small model sends each question to a code, math, research or chat agent via Ollama.
UPDATED · 03.10.2026RAG on GX10: Answers Citing Your Documents
Local RAG on GX10: Ollama for embeddings and answers, Chroma for vectors, answers with citations and a refusal when the documents lack the answer.
UPDATED · 03.10.2026Text to Knowledge Graph with a Local LLM
From text to a knowledge graph on GX10: a local model extracts entities and relations, Neo4j stores them, and you ask in plain language.
UPDATED · 03.10.2026Semantic Video Search on GX10 with Local AI
Find a scene in a video with one sentence: frames, descriptions from a local vision model, embeddings and ChromaDB search with Gradio on GX10. GDPR notes.
UPDATED · 03.10.2026OpenClaw on GX10: a Personal AI Agent on Your Own Machine
Install OpenClaw on an NVIDIA GB10-class server: Gateway as a service, dashboard over an SSH tunnel, a first approved task and sandboxing.
UPDATED · 03.10.2026NemoClaw: OpenClaw in a Sandbox on GX10
NemoClaw is an NVIDIA reference stack that runs agents like OpenClaw in an isolated OpenShell sandbox: requirements, install and a first prompt on GB10.
UPDATED · 03.10.2026A Shell Command Agent with Approval and Safeguards
A shell agent with human approval: allowlist, no shell=True, timeout and audit log, using a local Ollama model, plus the role of NVIDIA OpenShell.
UPDATED · 03.10.2026Live VLM: Live Vision with a Local Model
A local model describes what it sees in a video or on a camera: OpenCV, Ollama and qwen3-vl, one question at a time, an SSH-tunnel web view and GDPR rules.
UPDATED · 03.10.2026CUDA-X Data Science: pandas and scikit-learn on the GPU on GX10
Speed up pandas and scikit-learn on the GPU of a GX10 (GB10) with cuDF and cuML: cudf.pandas, cuml.accel, profiler. Checked 3 Oct 2026.
UPDATED · 03.10.2026Portfolio Optimization with cuOpt and cuML on GX10
Mean-CVaR portfolio optimisation with cuML and cuOpt on a GX10 (GB10), per NVIDIA's playbook. For learning, not financial advice. Checked 3 Oct 2026.
UPDATED · 03.10.2026Single-cell RNA on GX10: GPU Analysis
Analyse single-cell RNA data (scRNA-seq) with RAPIDS and rapids-singlecell on GB10: the NVIDIA playbook, public datasets, limits. Not medical advice.
UPDATED · 03.10.2026JAX on GB10: jit, grad and vmap in Practice
Install JAX with CUDA 13 on a GB10 (ARM64), check that it runs on the GPU and learn jit, grad and vmap with a small neural network. Official sources only.
UPDATED · 03.10.2026Monitoring a GB10 Machine: the Built-in Dashboard, nvidia-smi and nvitop
How to watch load, memory and temperature on an NVIDIA GB10-class machine: the built-in DGX Dashboard, nvidia-smi, nvitop and a small page of your own.
UPDATED · 03.10.2026Tailscale: Reach Your GX10 from Anywhere
Connect an NVIDIA GB10-class machine to your devices with Tailscale: SSH from anywhere, services only inside your network, access rules and key lifetimes.
UPDATED · 03.10.2026VS Code Remote SSH: Work on a GX10 from Your Own Computer
How to write and run code on an NVIDIA GB10-class server from your own computer: an SSH key, VS Code Remote - SSH, extensions and safe port tunnels.
UPDATED · 03.10.2026Two Sparks: a Two-Agent Capstone on GB10
A capstone: an analyst agent (RAG with citations) and an executor agent (n8n with human approval) on a local AI server. On one machine or two.
UPDATED · 03.10.2026Three GB10 Machines: When to Choose a Ring
Three GB10 machines: a ring, a switch or one cable? A short lesson on the choice per the NVIDIA guides as of 03.10.2026; the steps are in lesson 04-204.
UPDATED · 03.10.2026A GB10 Cluster Through a QSFP Switch
Four or more GB10 machines through one QSFP switch: 200 Gbps speed, a bridge, addresses, SSH and an NCCL test, following the NVIDIA guide as of 03.10.2026.
UPDATED · 03.10.2026NCCL on 2–4 GB10 Machines: Testing the Link
We build NCCL and nccl-tests on two to four machines of the NVIDIA GB10 class and run a bandwidth test of the link following the NVIDIA playbook.
UPDATED · 03.10.2026Isaac Sim and Isaac Lab on GB10: Robotics in Simulation
Is Isaac Sim supported on GB10 (ARM64)? Yes, with limits: build from source, Isaac Lab, a first training run, following NVIDIA's official documents.
UPDATED · 03.10.2026Reachy Mini and GB10: A Photo Booth with a Local AI
Run NVIDIA's "Spark & Reachy Mini" reference project on a GB10-class machine: a robot that talks, shoots and restyles photos. Licence and personal data.
UPDATED · 03.10.2026FourCastNet on GB10: AI Weather Forecasting with Earth2Studio
Run FourCastNet 3 with the Earth2Studio library on a GB10-class machine: data, forecast, map. Not an official forecast — those come from NIMH.
UPDATED · 03.10.2026CorrDiff on GB10: A Finer Regional Forecast
How CorrDiff sharpens a coarse global forecast into a regional field on GB10 with Earth2Studio: models, memory, modes. Not official — those come from NIMH.
UPDATED · 03.10.2026Video Search with Cosmos-Embed1 on GX10
Search a video collection with a sentence: Cosmos-Embed1 turns clips and text into vectors, Qdrant in Docker ranks them. A lesson for GX10 (NVIDIA GB10).
UPDATED · 03.10.2026Cosmos Predict 2.5 on GX10: Video from One Image
NVIDIA Cosmos-Predict2.5: from one image and a prompt to a short video on GX10 (GB10). Install with uv, licence and an honest list of unknowns.
UPDATED · 03.10.2026Cosmos Reason2 on GX10: Video Reasoning
Cosmos-Reason2-8B on GX10 (GB10): questions about video with step-by-step reasoning, Transformers and vLLM, licence and the model's limits.
UPDATED · 03.10.2026PINNs for Flow: a Digital Twin with PhysicsNeMo on GX10
How physics enters a neural network's loss: the official PhysicsNeMo example and a small PyTorch PINN that finds an unknown force from sensors, on GX10.
UPDATED · 03.10.2026Route Planning with cuOpt on GX10
How to model a routing problem (vehicles, stops, capacity, time windows) and solve it with NVIDIA cuOpt on GX10 — honest measuring, no promised numbers.
UPDATED · 03.10.2026A Foundation Model for Transactions: Embeddings from Sequences
How transactions become “words” for a model and embeddings become features: a small teaching model on invented data and an honest look at NVIDIA’s example.
UPDATED · 03.10.2026Fraud Detection with Graphs and XGBoost
A lesson on card-fraud detection: a transaction graph, XGBoost, AUC-PR and explaining a score. Honest about NVIDIA's hardware needs. Not financial advice.
UPDATED · 03.10.2026Portfolio Optimisation with Constraints and Factor Risk
A sequel to Mean-CVaR: a factor risk model for many assets and constrained optimisation, with a cuOpt example. An educational lesson, not financial advice.
UPDATED · 03.10.2026Distilling a Language Model for Financial Text
A large model teaches a small one to classify financial headlines: teacher labels, LoRA, F1 scoring. Honest on memory and hardware. Not financial advice.
UPDATED · 03.10.2026Deepfake Image Detection: A Service with a Human in the Loop
A service that asks a detector if an image looks synthetic, keeps only a hash and the result, and sends doubtful cases to a person. As of 03.10.2026.
UPDATED · 03.10.2026Digital Human on GX10: A Voice Character with Local Models
A voice character with Parakeet, Nemotron Mini through Ollama and Piper: model licences, English-only limits and AI disclosure. As of 03.10.2026.
UPDATED · 03.10.2026Audio2Face-3D on GX10: Facial Animation from Voice
Audio2Face-3D turns speech into facial animation. What the docs say: container 2.0, licences, GB10 limits and avatar rules. As of 03.10.2026.
UPDATED · 03.10.2026VISTA-3D: CT Segmentation with a Local Model
How NVIDIA VISTA-3D works: organ segmentation from CT, requirements, licence and limits, public data and checking the result. Not for clinical use.
UPDATED · 03.10.2026A Biomedical Literature Agent: PubMed and PDB
A local-model agent: searches PubMed and PDB, writes a cited report, checks citation numbers. NVIDIA's AI-Q blueprint is deprecated. Not medical advice.
UPDATED · 03.10.2026AlphaFold2: Protein Structure and pLDDT
How AlphaFold2 predicts a protein's shape, what the NIM container needs in disk and CPU, how to read pLDDT and which licences apply. Not for clinical use.
UPDATED · 03.10.2026Genomic Analysis with Parabricks on GX10: From Reads to Variants
GPU-accelerated genomic analysis with NVIDIA Parabricks on GX10: from FASTQ through fq2bam to VCF with haplotypecaller. Public data only, no clinical use.
UPDATED · 03.10.2026Molecular Docking with DiffDock NIM on GX10
How to predict how a small molecule binds to a protein: DiffDock as an NVIDIA NIM in Docker, public data and an honest answer on GB10 compatibility.
UPDATED · 03.10.2026Evo 2 on GX10: A DNA Model — What Fits and What Does Not
What Evo 2 is, what fits on a GX10 (a memory calculation) and how to generate DNA via the API and with the 7B model. Educational examples only.
UPDATED · 03.10.2026OpenUSD Digital Twin on GB10: Layers, Variants and Checks
Digital twin with OpenUSD on a GB10-class machine (ARM64): build from source, layers and variants, a validator, and an honest look at USD Code and Search.
UPDATED · 03.10.2026Robot Fleets: Mega and Routing with cuOpt on GB10
How robot fleets are tested in a digital twin (NVIDIA Mega) and how to plan routes with cuOpt on GB10 (ARM64), with an honest map of what is supported.
UPDATED · 03.10.2026Synthetic Robot Motion: Isaac Lab Mimic on GB10
From 10 human demonstrations to a thousand: Isaac Lab Mimic for synthetic robot motion on GB10 (ARM64) — steps, honest figures and limits.
UPDATED · 03.10.2026Isaac GR00T N1.7 on GX10: A Model for Humanoid Robots
Isaac GR00T N1.7 is an open vision-language-action model for robots: setup on GB10 (DGX Spark), a dry run, fine-tuning and limits. As of 03.10.2026.
UPDATED · 03.10.2026Video Agent with NVIDIA VSS on GX10
Stand up NVIDIA's ready-made VSS video agent: upload a recording, ask in plain language, get a report. Checked against the docs as of 03.10.2026.
UPDATED · 03.10.2026Nemotron 3 Nano Omni on GX10: Video, Audio and Images
Run Nemotron 3 Nano Omni (31B, ~3B active) with vLLM in Docker on a GX10: video, audio, image and text in, text out, all local, with a closed port.
UPDATED · 03.10.2026Nemotron OCR on GX10: Text from Images, Locally
Run NVIDIA's Nemotron OCR v2 as a container on GB10: a request, a response with coordinates, a Python client and field extraction with a local model.
UPDATED · 03.10.2026Detecting Synthetic Video with NVIDIA NIM on GX10
Run NVIDIA's synthetic video detector on your own server: what it returns, how far it is from proof, and why a person decides. Checked as of 03.10.2026.
UPDATED · 03.10.2026Nemotron 3.5 Content Safety on GX10: a Guardrail for Your Chatbot
A local 4B safety model checks the question, the image and the chatbot's reply: safe or unsafe, optionally by your own policy. vLLM, closed port.
UPDATED · 03.10.2026Active Speaker Detection on GX10: Who Is Speaking in the Video
The Active Speaker Detection NIM tells who is speaking in a video. Launch, inputs, sample client, limits and personal-data rules. As of 03.10.2026.
UPDATED · 03.10.2026LipSync NIM on GX10: Lips Matched to New Audio
LipSync NIM puts the lips in a video in step with new audio. Approved access, gRPC, languages, licences and consent per the docs as of 03.10.2026.
UPDATED · 03.10.2026AI Relighting on GX10: New Light on the People in a Video
NVIDIA's Relighting NIM changes the lighting on people in video from an HDR map. Launch, client, settings, consent and AI disclosure. As of 03.10.2026.
UPDATED · 03.10.2026FLUX.2 [klein] on GX10: Fast Images, Editing, and Which Variant Is Fine for Paid Work
Run FLUX.2 [klein] on a GX10 (NVIDIA GB10, ARM64) with ComfyUI and diffusers, and check the licences: 4B is Apache-2.0, 9B is non-commercial only.
UPDATED · 03.10.2026DeepSeek V4 Flash and GX10: the sums, and access through the API
Does DeepSeek V4 Flash (284B) fit in the 128 GB of memory of a GX10? The arithmetic from official pages, API access and a fenced coding agent.
UPDATED · 03.10.2026Kimi K2.6 and GX10: why it does not fit, and how to use it through the API
Does Kimi K2.6 (1T parameters) fit in the 128 GB of memory of a GX10? The arithmetic from official pages, API access and what leaves the machine.
UPDATED · 03.10.2026Local RAG on GX10: Vectors, Reranking and Answers with Citations from Your Documents
Build a local RAG on a GX10 (NVIDIA GB10, ARM64) with NeMo Retriever for vectors and reranking, Milvus Lite and a local model that cites sources.
UPDATED · 03.10.2026Scanning Your Own Containers: Trivy and a Local AI for Priorities
Scan your own Docker images on a GX10 with Trivy, rank the findings with a local model and keep an SBOM. Defensive use only, no exploits.
UPDATED · 03.10.2026Agentic Commerce on GX10: ACP, UCP and a Safe Checkout
How ACP and UCP work, what the NVIDIA blueprint requires, and how to build a checkout where an AI agent prepares the order and only a human pays.
UPDATED · 03.10.2026Confidential RAG: Encryption in Use and Its Limits
How to set up RAG for sensitive documents: tenant isolation, TLS, an audit log, encryption in use (TEE) and where its limits lie, on a GX10.
UPDATED · 03.10.2026Research Agent on GX10: Plan, Search, Cited Report
A small agent on a local AI server: plans sub-questions, searches your documents and writes a cited report. Checked against NVIDIA AI-Q, 03.10.2026.
UPDATED · 03.10.2026Voice Agent on GX10: Nemotron in Real Time
Run NVIDIA's Nemotron Voice Agent on a GB10-class machine: ASR, LLM and TTS via Docker Compose, the DGX Spark profile, languages and limits.
UPDATED · 03.10.2026Data Flywheel on GX10: Learning from Real Traffic
How real traffic to a model becomes a better model: logging, selection, comparison and a human decision. NVIDIA's blueprint is deprecated (03.10.2026).
UPDATED · 03.10.2026Catalogue Enrichment on GX10: From a Photo to a Reviewed Record
How the NVIDIA catalogue blueprint works, what fits a GB10 machine, and how to build a pipeline in which a human approves every draft record.
UPDATED · 03.10.2026Shopping Assistant on GX10: What Fits and How to Build It
What the NVIDIA shopping-assistant blueprint contains, why it does not fit one GB10 machine, and how to build a smaller assistant from parts that fit.
UPDATED · 03.10.2026Draft a Visit Note from a Consultation with Local Models
A local model turns a recorded consultation into a draft note and the doctor approves it. NVIDIA's blueprint is not for GX10. Not medical advice.
UPDATED · 03.10.2026Video Dubbing with Lip-Sync on GX10
Translate and dub a video on a local GB10-class AI server: transcription, translation, voice synthesis and lip-sync. Honest about licences and limits.
UPDATED · 03.10.2026A Multi-Agent Warehouse: Proposals That a Human Approves
A small system with roles for stock, forecast and anomalies in which a human approves every proposal. NVIDIA's blueprint recommends 4 × H100 to self-host.
UPDATED · 03.10.2026Your Own Model in a NIM Container on GX10
How to run your own or fine-tuned Hugging Face model in a Multi-LLM NIM on GX10 with an OpenAI-compatible API: keys, memory, security and checks.
UPDATED · 03.10.2026Streaming Data to RAG on GX10: Live Search
How to feed a stream of events (Kafka) into a Milvus vector database and query the live index on a local GB10 server. What NVIDIA’s example is and is not.
UPDATED · 03.10.2026AI Factory Digital Twin: DSX and OpenUSD
What Omniverse DSX is, what it needs (RTX Pro 6000, not GX10), and how to build a small twin of a server room in a text USD file with a power budget.
UPDATED · 03.10.2026Flood Risk: Runoff and a Rough Flooded-Area Estimate from a DEM
A teaching lesson: SCS-CN runoff and a rough flooded-area estimate on the Copernicus DEM in Python. Not a warning; official ones: NIMH and GDPBZN.
UPDATED · 03.10.2026Syncthing: Direction Is Not Preservation
A Syncthing "Send Only" folder does not protect an archive from deletions. How ignoreDelete and Trash Can versioning work and how to set them up.
UPDATED · 03.10.2026New Model, Old Runtime: How to Test Safely
A new model returns HTTP 500 at once? Often the runtime is too old for its architecture. How to spot it in the log and test a newer one on another port.
UPDATED · 03.10.2026The Heavy Model Without a Crash: Context Is the Culprit
A large context can overflow memory even for a small model. How to size the KV cache and set context, Flash Attention, q8_0 and keep_alive in Ollama.
UPDATED · 03.10.2026Service vs Hand-Started Process: Arm Everything You Rely On
Why one component comes back by itself after a crash while another stays down: a systemd service with Restart=on-failure versus a hand-started process.
UPDATED · 03.10.2026Games on an ARM Machine: Three Layers of Translation
How x86 Windows games reach an ARM64 Linux machine: the FEX, Proton and remote-stream layers, what each one costs and why the bottleneck moves.
UPDATED · 03.10.2026Model Broker: Keeping Shared Memory from Crashing the Server
How to keep a unified-memory AI server alive: the context as a hidden bomb, a fence in Ollama and a budget-queue broker. Checked against docs 03.10.2026.
UPDATED · 03.10.2026Safe One-Way Synchronisation: Syncthing That Does Not Delete
How to sync a computer to a server in one direction without losing the second copy: folder roles, trash can, ignoreDelete and a test. Checked 03.10.2026.
TESTED · 3 Oct 2026UPDATED · 03.10.2026File Deduplication with an Audit Trail: Proof First, Then Deletion
Find duplicates without risk: name and size as a candidate, BLAKE3 as proof, czkawka_cli and a log written before deletion. Checked 03.10.2026.
VERIFIED · 01.10.2026UPDATED · 01.10.2026Cyrillic Paths in PowerShell and ssh: the Encoding Trap
Why Cyrillic paths break between Windows PowerShell and ssh, how to catch it (code pages, UTF-8, BOM) and how base64 carries a path without any loss.
UPDATED · 03.10.2026Which AI Models Fit on a GX10: Size, Context and License Compared
Compare open AI models for GX10: file size, context, input type and license from public model pages, dated, with a rule for the 128 GB memory.
UPDATED · 03.10.2026How to Measure the Speed of an AI Model on a GX10 Yourself
Learn to measure the speed of a language model on a GX10 yourself: conditions, the Ollama API, repeats, the median and a measurement card.