faster inference
Shelf photo end-to-end: 17 s → 0.7 s after the GPU migration.
serving cost
Inference service rewritten in Rust + ONNX Runtime: $66.70 → $13.71 / month.
SKU accuracy
Product classifier measured on 31,311 unseen photos.
errors / 8,833 runs
Seven days of production traffic, all served on GPU.
01 / production machine learning
Numbers straight from the production logs.
response time in production · last 7 days
typical slowest 5%8,833 runs · 0 errors · peak 3,287 runs/day
02 / model lab
What each model delivers today: how many brands it recognizes, how accurately, how fast, and how much it saves.
Identifies the brand, sub-brand and exact SKU of every product on a shelf.
Added 7 new SKUs at 97% accuracy with no new labeling round — ~$800 saved per extension.
Tells which of 11 store types each visit belongs to, from its photos.
Three of every four visits are classified with no human review; only doubtful ones reach a person.
Recognizes the cold equipment in a store and which brand owns it.
Brand-level cooler audits answered in 25 ms per photo, live in production.
Finds signage, fridges, shelving, entrances and address plates in store photos.
Replaced a paid LLM step: up to 160× faster, at about one cent per thousand photos.
Locates every product on a crowded shelf so each one can be recognized.
Turns one shelf photo into ~41 recognized products in under a second.
Spots shops in street-level video to map new points of sale.
Annotators start from the model's boxes instead of empty frames.
03 / ai agents in operations
Two agent systems I built and run: one answers business questions from live data, the other runs field work on WhatsApp.
5.8k research questions + 806k field-capture messages.
LLM calls, data queries, photo checks, transcriptions.
Captured through the agents and checked by AI on arrival.
Field teams on WhatsApp, ops and sales on Slack.
research-agent questions / month
* October through the 21st
LLM agents on Vertex AI · custom MCP server · WhatsApp Cloud API · speech-to-text · BigQuery · PostgreSQL/PostGIS · Pub/Sub · Slack
04 / more work
Real-estate intelligence I founded and personal projects to test my agents and manage teams of developers and AI agents. More below: Risks Vision quant research, a city-scale retail census from my current role and an on-chain game economy.
Every listing gets a model-estimated fair value, so buyers instantly see what is under- or over-priced.
Each listing is checked before buyers see it — inconsistent or suspicious ones never reach them.
The real production model, running in your browser. Change any field and the estimate updates instantly.
Estimated value
—loading model…
Model accuracy on unseen listings: R² — · avg. error —
Which LLM is really better for a task? I don't guess. I put the models in the ring with the same test cases and score every round: accuracy, cost, speed and reliability. Any model can enter, and every fight ends with what to improve.
Every model gets the exact same test cases.
One round per metric. Any model can enter: JEV, Gemini, Claude, GPT, DeepSeek…
Who wins, where each one loses, and what to improve before shipping.
| Round | JEV | Gemini | Winner |
|---|---|---|---|
| R1 Accuracy who answers right more often | 94% | 88% | JEV · +6 pts |
| R2 Cost per 1k queries · cheaper wins | $0.17 | $0.90 | JEV · 5× cheaper |
| R3 Speed p95 latency · faster wins | 0.85s | 2.4s | JEV · 3× faster |
| R4 Reliability F1 across every class | 0.92 | 0.85 | JEV · +0.07 F1 |
Next step → Gemini costs 5× more and is 3× slower here, for less accuracy. Not the model for this task.
Who ships, how fast, and whether the code holds up, for developers and for the AI agents writing code next to them. Below is a real run on a public open-source repo, so every number can be checked.
who writes the code
management report · 27 Jul – 24 Oct 2026 · 90 days
👤 without AI🤖 with AI
→ AI code is redone 1.6× more. Review AI PRs more carefully before merge, then measure again in a month.
→ config/scripts depends on one person (66% of its files). Pair the next contributor on its reviews.
→ Team 4 people make 80% of the changes. Succession plan: document what they know, spread the reviews.
Check it yourself: github.com/stablyai/orca ↗ · merged PRs and main-branch history, 27 Jul – 24 Oct 2026 · 90 days. People are anonymized.
Machine-learning strategies, run by AI agents: one reads markets, one sets risk limits, one runs the book. 102 strategies were stress-tested with real fees — only 8 made the portfolio.
Out-of-sample (2025+): profit factor 1.32 over 612 trades · beta to the S&P 500: 0.02 · simulated results, not live returns.
A living census of traditional retail: where every store is, how much potential each zone holds, and where field teams should go next.
Every reward, trade and stake in this fantasy RPG settles on-chain through smart contracts I wrote.
05 / experience & skills
24× faster inference · 812k+ agent messages
18.5k listings monitored · 97.8% best audit ROC-AUC
102 strategies tested · 33% CAGR backtest
100k+ on-chain transactions · Go & Solidity
contact / next collaboration
Tell me about the model, agent or data system you need. Email is the fastest path — the form drafts the message for you.