House Numbers
Core engineer on a 16-service TypeScript monorepo for an AI-powered mortgage-lending platform, working across backend services, frontend, and infrastructure. Joined when the engineering team was three people; the team has since grown to seven engineers, excluding design, PM and QA.
- Built and still owns the platform's communication service end to end: asynchronous, webhook-driven email and SMS/call intake, LLM-assisted message classification kept separate from the deterministic rules that route each message, and loan matching over field-level-encrypted data.
- Redesigned and rebuilt the security boundary of the platform's public document-upload flow after an estimated two in three submissions had ended in a borrower complaint or support escalation: an opaque public session, CAPTCHA and rate limiting, replay protection, and a hybrid deterministic/LLM routing model that kept ambiguous matches with a human, all covered by a correlated audit trail.
- Led the multi-month migration of the platform's LLM infrastructure off a prompt-management vendor (Orq.ai) across four services with no significant regression: version-controlled prompts, provider failover, per-call cost tracing, and evaluation harnesses combining human-graded ground truth with LLM-as-judge scoring. Designed the *.agents.ts convention that isolates model-dependent tests into a CI lane triggered only when a prompt or agent changes, and moved model selection out of environment variables so judging can use a provider other than the one being judged. Also replaced hallucination-prone LLM-generated document checks with deterministic, tested validation.
- Built and evolved flexible query filtering for the loan-application product end to end: the frontend filter flows and a reusable TypeScript filter library and query parser that compose MongoDB queries from them — multi-value, related-entity, date-range, set/unset, and regex-escaped text-search operators — with tenant-scoped access for sensitive loan data.
- Designed and delivered a monorepo-wide, agent-consumable documentation system with deterministic integrity gates in CI, alongside broad ownership of shared packages, public upload flows, user offboarding, observability, and production incident response.
- Led a multi-round proof of concept evaluating vision-LLM document splitting, classification, and extraction as a candidate replacement for the incumbent OCR vendor, including the human-reviewed corpus, evaluation and cost tracing, and fixes to measurement bugs in the scoring. The PoC stayed experimental: production kept the vendor plus the existing LLM fallback, and productionization was left as a later team decision.
- typescript
- nodejs
- react
- mongodb
- aws
- ecs
- redis
- kubernetes
- openai-api
- anthropic-api
- new-relic
- zod