Case Studies
Case Studies
How each system works, the constraints and tradeoffs behind it, and the parts that didn't work the first time.
Case StudyNexus AI
Cites the exact line and scores options with calibrated confidence
A LangGraph assistant that shows its route, cites the exact page and line from your documents, and scores decisions with a calibrated decision model. Built on FastAPI, Next.js and JEV.
- 18-cell decision board scored
- 0.4s
- Faster backend cold start
- ~2×
- LLM providers supported
- 8
Case StudyMLInsights
68 ML models, no code, and a backend that scales to zero
I rebuilt MLInsights so anyone can upload a spreadsheet, then train, compare and export 68 ML models in the browser. The backend has no database and scales to zero, which forced careful choices about storage, long-running training and, unexpectedly, bot traffic.
- ML models available
- 68
- Comparison views
- 7
- Database servers
- 0
Case StudySelf-Driven Car
Real neuroevolution in the browser, with the actual training code stepping live
I rebuilt a Pygame neuroevolution project as a browser lab where anyone can watch cars learn to drive with NEAT. A vectorised NumPy engine streams live frames over WebSockets, and the code panel shows the real Python source running each step.
- Cars per generation
- 20–150
- Backend tests
- 26
- Accounts needed
- 0
Case StudyGrabbit
A media downloader that never offers a format it cannot deliver
Grabbit turns a public link from 15 sources into exactly the file you want, offering only the qualities and formats the media and server can genuinely produce. I built it with a FastAPI, yt-dlp and ffmpeg backend and a static Next.js frontend.
- Backend tests
- 126
- Supported sources
- 15
- Cold start
- 0.45–0.71s
Case StudyLogvera
Five anomaly models vote on every log window and explain themselves
A private log analyser that turns any server, network or app log into a health score, explained anomalies and security findings. Five ML detectors vote on each time window, and an LLM summarises only redacted aggregates.
- Anomaly models voting
- 5
- Upload retention
- 60 min
- Backend tests passing
- 41
Case StudyComet
A link shortener that blocks phishing before the link exists
Comet is a URL shortener where every link is checked by rules and an AI model before it exists, encrypted at rest, and deleted within 7 days. I built it to run on one small container, with a hand-written Bloom filter, HyperLogLog and request tracing.
- Backend tests passing
- 70
- Cached lookup median
- 7µs
- Visitor IPs stored
- 0
Case StudyDecideX
Catches 85.6% of leavers on held-out data, then drafts the offer
I built a churn decision app that reads account notes with an LLM, scores customers with a calibrated model and recommends what to do, with reasons. A cost-tuned threshold catches 85.6% of leavers on held-out data.
- Churners caught
- 85.6%
- Missed churners
- 184→54
- Decision time
- ~1.5s