Projects

Systems, not demos.

I build things that resemble production: real architecture, real constraints, and AI used where it earns its place. Here's what I'm working on.

A production-grade, AI-assisted, multi-tenant SaaS CRM.

Problem: Sales and account teams drown in manual CRM upkeep. Emails come in, context gets lost, and follow-ups slip. I wanted CRM software that reads the inbox, proposes the next action, and keeps a human firmly in control.
Technical depth: Full-stack TypeScript + Python platform with multi-tenant workspace isolation, Gmail OAuth inbox sync, AI-assisted email analysis, and a human-in-the-loop approval flow whose execution engine resolves dependencies and writes a linked CRM graph. Stripe billing, RBAC, CSRF protection, JWT/session auth, background workers, and job queues — containerized for Docker, designed for Kubernetes. Core CRM, AI workflows, workspace isolation, auth, billing, and inbox sync are operational; current work is deployment, observability, and production hardening ahead of a small public beta.
What I learned: How to keep AI useful but safe behind an explicit approval boundary, how to design tenant isolation you can actually trust, and how much of 'production' is the unglamorous work: migrations, auth edge cases, and deployment.

In development

AI-assisted construction takeoff platform.

Problem: Construction estimating is slow, manual, and inconsistent. Plan review and quantity takeoff eat hours of skilled time and live in a few experienced people's heads. I'm building a platform that performs plan analysis, understands construction documents, and extracts measurements — designed around how estimators actually work.
What I learned: Domain-specific AI lives or dies on data structuring and trust. The interesting problems are upstream of the model: clean inputs, clear review surfaces, and never letting a confident wrong number through unchecked.

PythonFastAPIComputer visionDocument understandingMeasurement extractionPostgreSQL

AI-assisted decision systems for quantitative finance.

Problem: How should humans and AI collaborate on high-stakes financial decisions? I'm interested in decision systems where models surface signal and structure, but accountability stays with people.
What I learned: Framing the research question well matters more than the model. The hard part is evaluation: defining what 'good' means for a decision system humans actually rely on.

More is in progress.

I'm continuously building. If you'd like a walkthrough of any project — architecture, code, or decisions — reach out.