Case studies
Work we have shipped.
Projects from 2024 to 2026, all running in production. Each case study covers the problem, what we built, the decisions behind it and what went wrong along the way.
10 projects

92% F1 score
Search over company documents
Answers questions from handbooks, wikis and PDF archives, with a citation for every claim. Hybrid retrieval and an evaluation gate in CI keep the quality from slipping.

71% of tier-1 tickets resolved without staff
Bilingual support agent
A LINE and web chat agent for a Japanese retailer. It answers routine tickets in Japanese or English and passes the rest to staff with a short summary, so customers never repeat themselves.

99.1% defect recall
Visual inspection on a factory line
Defect detection for a Japanese manufacturer. Inference runs at the edge in 38 ms per frame, and operators can see why each part was flagged.

7 specialized sub-agents
Executive assistant agent
An assistant that works across Slack, Google Calendar, GitHub and email. People ask for things in plain language. It plans the steps and asks before doing anything risky.

Transcripts in Japanese and English
Bilingual meeting notes
Records meetings in Japanese and English, writes a summary in both languages, and posts action items with owners to Slack and Notion within two minutes of the call ending.

85% of SDR tasks automated
Sales pipeline automation
Qualifies inbound leads, updates the CRM, books demos and runs follow-up emails. It now handles about 85% of the work an SDR used to do by hand.

10k+ model calls a day
LLM monitoring dashboard
Tracks accuracy, latency, cost and toxicity for more than 10k model calls a day, and stops a deploy when evaluation scores drop.

87% bug catch rate
Pull request review agents
Paste a pull request link and three agents review it. One plans the review, one looks for bugs, and one writes fixes with tests and leaves comments.

85% pass@1 on HumanEval
Offline coding assistant
A VS Code assistant that runs on the developer's laptop with Qwen 2.5, Llama 3.3 or Mistral Small. No API bills, and the code never leaves the machine.

Charts update every 15 seconds
Trading terminal with AI signals
A trading dashboard with live candles, more than 25 technical indicators, model-generated signals, paper trading and backtesting.