A few projects we've shipped with forward-thinking teams. We can only share what's public — the most interesting work usually isn't.
Brex needed to categorise millions of transactions and reconcile in real time. We shipped an LLM-powered categoriser with eval-driven guardrails — and a 60% drop in manual review.
Replaced a brittle forecasting stack with a deep-learning pipeline tuned to local demand spikes. Less waste, fewer stockouts, calmer ops teams.
We replaced a queue of manual reviewers with a multi-modal pipeline that reads receipts, infers policy and writes audit trails — with humans in the loop where it matters.
Built a recommender that surfaces high-fit investors for founders and vice versa, with explanations a human can defend.
Most engagements start with a 30-minute conversation and a short architecture sketch. No SOW theatre, no boilerplate decks.
