The handover machine
Gemini reads a construction handover pack — operation manuals, drawings — and four regulatory deliverables come out the other side, unattended, in about twenty-seven minutes, with provenance on every field. The extraction is schema-constrained so the same document gives the same answer every run, which is what let a language model near documents that get certified at all.
Cloud Run · Pub/Sub · Gemini · JSON Schema · Pydantic
Insurers who disagree, reconciled
Ten-plus carriers, each with their own idea of what a policy record looks like, normalised into one standard shape for a US insurance brokerage. A rules engine and fuzzy matching behind the model, tiered so a person only ever reviews the genuinely ambiguous middle. A thousand policies, live in production, and the daily manual pass stopped.
Python · Flask · PostgreSQL · Redis
Cricket, quantified
Real-time extraction from live sources, a processing pipeline, predictive models and decision dashboards for Decidr Sports in Sydney — then a natural-language agent over the warehouse so coaches could ask questions in plain English instead of filing tickets. Fifteen months and four hundred hours, closed at five stars.
PySpark · AWS · Docker
The $847K correction
A quantity-surveyor solver validated to the cent against three separate workbooks. Its first run on a real building found the error.
Python · golden-file tests · pytest
Leads at a tenth of the price
Walmart scraping plus Apollo.io enrichment on a serverless schedule, replacing a manual workflow entirely — eighty per cent less cost at ten times the coverage.
AWS Lambda · S3 · Apollo.io
Scrapers that survive handoff
LinkedIn extraction at volume with anti-bot and CAPTCHA strategy, a bot reading structured data off a live video stream, and an arbitrage bot with real money on the line. All scheduled, monitored, and still running after I handed them over.
Scrapy · Selenium · Playwright · OpenCV