Compliance engines inside banks, multi-agent platforms in production, and the test engineering underneath software that gets audited. Selected work.
A compliance team spent hours converting Currency Transaction Reports into goAML XML by hand. We built the full conversion pipeline — parser, XML builder, schema validator, audit trail — behind a web UI officers actually use.
A registry of every periodic return a bank owes. The engine pulls the required data from its sources, builds each report against the mandated template, then schedules, prepares and submits it — behind a Compliance maker–checker step.
Law-enforcement requests arrive as letters with scanned attachments. The agent watches the mailbox, OCRs every attachment, extracts the national ID numbers and verifies them against the core banking system. An LLM classifier decides the intent — provide records, block, or unblock — and routes accordingly. Every step lands in a review portal with a full audit trail.
Watchlist alerts arrive continuously and most are false matches. The agent collects them, dispositions the clear ones against a rule engine backed by an LLM, and routes every disagreement to a human with the evidence already assembled. A read-only dashboard and a branded daily report keep Compliance across it without anyone chasing a queue.
An employee asks a policy question in plain English and gets an answer drawn strictly from the bank's approved documents — never from the model's own memory. A retrieval-augmented pipeline running behind a chain of guardrails.
Two agents on one platform, and a human approves everything before it ships. The newsletter agent researches what an industry is actually working through, clusters it into themes and writes a position on each — not a summary of other people's articles. The social agent plans a week of posts reviewed as a batch.
A genuinely autonomous agent running a custom ReAct loop: it researches trends, reads its own history to avoid repetition, writes a strategy memo, self-critiques its drafts, publishes, and learns from every rejection.
A microfinance bank ran IT, Operations, Compliance and Admin service requests over fragmented email. We replaced it with one platform — JSON policy engine, vendor module, department analytics — built with AI-assisted development.
A lending pipeline from application intake through risk scoring to disbursement, covered by a three-tier strategy: unit tests on the money logic, mobile automation for borrowers, BDD for the admin portal.
Surgical scheduling across hospitals, ASCs, surgeons, nurses and vendors. Manual regression was blocking weekly releases, so we automated the critical booking and alert workflows into CI.
An eClinical platform facing a HIPAA audit needed full automated evidence for its patient-reported outcomes module — across web, Android and iOS, on real devices.
A certification training platform for PMP, PgMP and ACP tracks. Admins build courses and a question bank with bulk CSV import, create timed mock exams from a separate pool, and manage time-limited enrolments. Students practise untimed, sit timed mocks and track readiness on a performance dashboard and study heatmap.
Most of these started as a single painful workflow someone described in a 30-minute call.