I build AI agents that have to earn the authority to act — not assume it.

An AI contract-intelligence engine for supply-chain exceptions: it reads contract amendments, resolves conflicting clauses by effective date and precedence, and prices penalty exposure in real time. Before I called it done, I built a chaos-injection suite to deliberately corrupt OCR numbers, plant contradictory amendments, and flood the review queue with 100 low-value events — specifically to see whether the ranking and precedence logic would survive an adversary, not just a demo dataset. When I did a pre-launch security pass on my own deployment, I found the approve/reject endpoints had no auth and open CORS — so I closed both before sharing the link, not after.

An invoice reconciliation agent for accounts-payable teams, built around one rule: the AI recommends, it never decides. A deterministic three-way match — pure Python, zero LLM — checks price, quantity, tax math, and duplicates against the PO and goods receipt. A Primary Agent classifies the invoice and explains why; an independent Critic Agent checks that reasoning against the match results and vendor risk history without ever overriding it. Every step writes to an append-only, hash-chained log that can be replayed without re-running the pipeline. Against 100 synthetic invoices with errors seeded across five categories, it caught all of them — while still auto-clearing 60% of clean invoices outright.

A month-end close reconciliation engine for GL-to-bank matching — the part of close that typically eats 6 to 10 business days. An Extractor agent normalizes messy statement data; an Auditor agent runs exact and fuzzy matching and flags anomalies; a Governor agent applies materiality thresholds and escalation logic and keeps an immutable audit trail. The LLM only ever explains — it never approves, rejects, or edits a record. Validated against 60 automated tests covering duplicates, timing differences, FX rounding, and missing entries: 97.5% match rate on a clean month, 88.8% on a deliberately messy one, 95.9% even on a seeded fraud scenario.

Turns SEC filings — 10-Ks, 8-Ks, earnings transcripts — into a one-page, board-ready strategic memo with a citation on every number, in minutes instead of the day an analyst would spend on it. A five-stage pipeline fetches EDGAR filings, extracts risk factors and MD&A with citation metadata attached at every stage, has Claude analyze variance and tone shifts year over year, and synthesizes a schema-validated memo. The last stage is a deterministic audit pass: it resolves every citation and checks every number against source — and if anything has drifted, it forces a re-synthesis before the memo ships, not after someone reads it.

A staffing-planning tool for appointment-based businesses that spends as much effort trying to break its own schedule as it does building it. It forecasts demand, sizes a staffing plan with a real safety cushion, and scores no-show risk per booking — then runs 528,000 simulated bad days to find exactly where that plan fails, and writes the recovery playbook in advance. Every number on screen is tagged MEASURED, ESTIMATED, or ASSUMPTION — it will not claim an operating history it doesn't have, and an internal auditor re-runs the production math on every plan before it's trusted.