AI on modest hardware · built with Claude Code
I help software teams run AI on their own GPU, or a rented one, with no per-request bill.
Seferehov reads whole bookshelves from a phone camera, using AI models on one 2016 graphics card in my home. Kartis7, a ticket-sales platform, checks every event page for scams with a model on that same card. Neither pays per request, so there is no bill that grows with traffic or with a stuck loop. I measure every change on real data, and a person approves the decisions that matter.
I have been building business software since 1993, and today I build with Claude Code every day. I advise teams on which models, which card and which checks they need, and I build the first version with them. Every week I publish a video and a post showing how, with real numbers.
New post
Seferehov: a night shift for Claude Code, and a camera for shaky hands
Four days after the last update: Claude Code now runs experiments on my home server every night, on my own subscription, locked into a user that can only read. The first one found 59 of 68 books on a test shelf instead of 47. And the camera no longer waits for a perfectly still hand.
Read the post →- Building software since 1993
- Claude Code every day
- Small models on an 8 GB GPU
- Your own GPU or a rented one
- Evals on every change
- Backend architect, CISOteria (7 yrs)
- Flixel founder · 7 patents
- B.Sc. Math & Physics, TAU
How I can help
Small open models can do many AI tasks well on a modest graphics card, if someone measures which model, on which card, for which task, and builds the checks around it. That is what I do.
Local AI Pilot
One AI feature running on your own GPU, or a rented one with a fixed monthly price, in 30 days. Measured against a cloud model on your own data, then handed over to your team. Fixed price, and the scoping week is refundable.
GPU and Claude Code review
Two hours on what your AI plans actually need: which models, which card, rent or buy, and how your team uses Claude Code. You get a written report with numbers.
Ongoing advice
A weekly session with your developers while they build: architecture, model choice, evals and guardrails. For teams already building with Claude Code who want a second pair of experienced eyes.
Proof: projects on one 8 GB GPU and beyond
Seferehov — reading bookshelves
Live · seferehov.com · vision + OCR on one 8 GB GPU
A map of free street libraries where a phone photo of a shelf becomes a list of books. A vision model finds the spines, a Hebrew OCR model and a spell-check read them (65% → 96% text accuracy), and a local copy of Open Library matches the titles. All on one 2016 graphics card.
Kartis7 — ticket sales
Demo · kartis7.com · AI scam screening
Ticket sales for small event producers, where the money goes straight to the producer. Every new event page is checked for scams by a model on my own GPU, and an assistant can manage events over MCP.
Small model beats Claude on BANKING77
Open source · LoRA fine-tune · English + Hebrew
A fine-tuned multilingual model sorts banking messages into 77 intents: 94.2% in English against 85.3% for Claude Sonnet, and 82.8% against 77.3% in Hebrew, in 12 ms instead of 2 seconds. Messages it is unsure about still go to the large model.
Compliance AI on the client’s own GPUs
Client work · DevOps & AI architecture
For a compliance software company: AI that runs on the client’s office GPUs with a fallback to CPU, so evidence never leaves their country. I built the deployment pipeline and review the architecture in weekly sprints.
RAG on Postgres / pgvector
Open source · 35M-chunk index · public demo
Search over all of Wikipedia with answers that cite their sources, on self-hosted Postgres and local models. 35 million chunks, about 1 ms per query. Live at ask.amiheines.com.
Guardrail — spend limits for AI
Live · LLM spend & abuse control
Hard spend caps in the request path, so a runaway agent loop stops at the gateway instead of on next month’s invoice.
More from me
Get new posts by email.
I write here occasionally — on AI-assisted coding and unusual, practical uses of AI. One email when a new post goes up, nothing else.
Or browse all posts →Have an AI feature in mind? Tell me what it should do. In 30 minutes I will tell you whether a local model can do it, roughly which card it needs, and what it would take to build.