The
builds.
Each entry is a working system I shipped — the problem, the decisions, the architecture, and what I'd change. Click through for the full case study.
Real-Time Sales-Conversation Coaching Agent.
How I'd build a real-time conversational coaching agent for in-person field sales (HVAC, roofing, home services). Two-path architecture: a live sub-2s feedback loop running on a streaming reasoning model, plus a post-call deep-analysis loop on a long-context model. Multi-speaker citation, per-tenant rubric, and the hardest tradeoff at the bottom.
MercuryStream.
Real-time market data pipeline with anomaly detection, schema-drift defense, and an in-memory black-box flight recorder that captures incident context only when something breaks.
AI Agent Error-Handling Patterns.
Four battle-tested reliability patterns for production AI agents — circuit breaker, partial success, human-in-the-loop, graceful degradation — implemented on Trigger.dev v4 with passing tests and copy-paste-ready upgrade paths.
Travel MCP Server.
An MCP (Model Context Protocol) server that gives AI assistants like Claude live access to flights, hotels, and weather. Backed by MongoDB caching and a fallback strategy for hostile data sources.
Jarvis.
Personal 'Life Operating System' built on Flutter + FastAPI. User data (conversations, memory graph, preferences) stays on-device; the backend is a thin stateless service for auth, LLM proxying, and billing only — it never sees real user data.
Apex Backtest.
Python wrapper around QuantConnect's Lean engine that hides the Docker / config / data-format complexity. Traders write a strategy file; the CLI handles environment, data downloads, container orchestration, and result formatting.
Structured AI · Founding Engineer.
Founding-engineer reps at an applied-AI company in NYC. Architected a production document-analysis agent on LangGraph + Anthropic Claude, an in-house code-execution sandbox for ZDR-compliant enterprise deployment, and an iOS / iPadOS client from zero. Most details NDA-bound — see the timeline modal for the public version.