← Home

Work

Selected projects.

LiveFull-Stack (React / Node.js / MongoDB)

Ledgic — AI-Powered Personal Finance Platform

Production finance app at ledgictech.com. Dual-token auth (HttpOnly cookies + Google OAuth 2.0), Claude API for AI-driven financial forecasts, and a 90-day spending prediction engine.

Highlights

Claude API (Anthropic)Dual-token JWT authAI financial forecastingServerless on Vercel

Artifacts

  • Live at ledgictech.com
  • 11-model MongoDB schema
  • Multi-turn LLM prompt engine
  • Full monorepo deployment
Prototype · YC applicantData Pipeline + Mobile (Snowflake / Claude / React Native)

Trauma Alert — Turning EMS radio chatter into real-time trauma alerts

A working prototype that ingests EMS radio chatter, transforms audio into structured data. Classifies each call with an LLM, and pushes real-time alerts to a mobile app built for emergency surgical teams. Runs end-to-end on mock audio clips through a Snowflake pipeline into a React Native client. Snowflake chosen to leverage platforms ML features, to evaluate and improve classifications over time, that schema is in place, though the MVP doesn't yet exercise it. Basis for a Y Combinator application.

Highlights

SnowflakeUnstructured DataLLM CategorizationReact Native

Artifacts

  • End-to-end pipeline on archived audio clips
  • Data classification
  • Real-time push to mobile client
  • Escalation logic for unacknowledged alerts
Internal Tool (Shipped)Multi-Agent LLM Pipeline (LangChain / Python)

AI Documentation Generator

An internal tool I built to cut the time my startup team spent writing and maintaining documentation. It analyzes a codebase and generates structured docs through a multi-agent LangChain pipeline, so new engineers could ramp on unfamiliar parts of the system faster and documentation stayed current instead of going stale.

Highlights

Multi-agent orchestration (LangChain)Codebase parsing & analysisAgent-to-agent QA guardrails

Artifacts

  • Four-stage agent pipeline: SWE → QA → Documentation → Final QA
  • Each stage validates the previous, catching low-quality output before it ships
  • Deployed internally to reduce manual documentation effort
  • Sped onboarding onto unfamiliar parts of the codebase