Projects

Side projects and applications I've built

BinaryBuilders

3 projects

Melio MealPlan AI

AI-Powered Meal Planning

**What it is.** Full-stack AI meal-planning platform built as a pnpm monorepo and deployed on AWS. NestJS owns business logic while a separate Python FastAPI ai-service runs the AI brain. **AI generation pipeline.** The AI brain is a single tool-using agent orchestrated as a LangGraph StateGraph with a ReAct inner loop that calls a USDA lookup tool but never writes nutrition numbers itself; the server computes every macro from USDA ground truth, a hard anti-hallucination guarantee. Two complementary RAG systems serve different jobs — a lexical USDA RAG (Postgres full-text + Haiku reranker) grounds every macro, while a separate semantic recipe RAG (HyDE → text-embedding-3-small → pgvector HNSW + full-text fused with Reciprocal Rank Fusion → Haiku reranker, seeded from RecipeNLG via a LlamaIndex ingestion pipeline) feeds curated dish ideas into a retrieve_recipes node before generation. **Reliability & quality.** Quality is enforced by a 3-layer validate-and-retry cascade, and typed state is checkpointed to Postgres for crash-resume. Validated meals flow back into the recipe corpus through a guarded data flywheel, an offline eval harness with a CI gate catches regressions, and every generation is traced in self-hosted Langfuse. **Deployment.** Server-Sent Events stream meals live from Python → NestJS → Next.js, all running on AWS — ECS Fargate plus Lambda/OpenNext over RDS, ElastiCache, and CloudFront.

NestJSFastAPIPythonPostgreSQLBullMQRedis+43 more
USDA Ground Truth
Server-computed macros — nutrition numbers can't be hallucinated
3-Layer Validation
Programmatic + LLM-judge + blocking USDA fact-check with retry

Feels Protocol

Solana DEX Trading Platform

Solana DEX trading platform built as a two-part system: feels-web, a Next.js 16 trading SPA owned end-to-end, and feels-indexer, a Rust real-time indexer over a four-backend storage layer. The frontend ships real-time market data, wallet-connected swaps, and on-chain position management across 30+ components; the indexer deserializes on-chain Solana account and event data and exposes it over a documented REST surface. A two-lane Playwright E2E setup (deterministic test wallet + LiteSVM in-memory RPC) keeps full-stack runs fast and flake-free.

Next.js 16React 19TypeScriptTailwind CSSBunRust+16 more
30+ Components
Full trading UX from market catalog to vault redemption
4 Backends
PostgreSQL + RocksDB + Redis + Tantivy storage layer

GovChime Analytics Platform

Government Contracts Intelligence

**What it is.** Government contracts intelligence platform over 4.2M federal contracts (PostgreSQL 16 OLTP) and 52M+ award rows (ClickHouse OLAP). **Data pipeline & ETL.** A custom ETL layer — three node-schedule-driven SmartSync services — ingests the SAM.gov Contract Awards API into PostgreSQL via idempotent composite-key upserts, materializes ~20 PostgreSQL materialized views, and mirrors the awards table into ClickHouse on a 6-hour cycle. Three overlapping sync layers plus a daily reconciliation job took SAM.gov restarts from ~3% record loss to zero data loss. **Analytics at scale.** The MV strategy (5 core pre-aggregations — agency, NAICS, geography, award type — covering ~80% of dashboard queries) cut p95 latency 25× (5s → <200ms) with no new infrastructure, while the ClickHouse columnar mirror serves no-filter analytics sub-50ms. **Deployment.** Next.js 15 ships on Cloudflare Workers (OpenNext) with ISR; the Express.js API and 3 SmartSync services run on Komodo bare-metal (Hetzner, Docker) behind 24+ GitHub Actions workflows on a self-hosted runner.

Next.js 15React 19TanStack Query 5TanStack Table 8Radix UIZod+32 more
25× Faster
p95 query latency 5s → <200ms via ~20 materialized views
4.2M Contracts
PostgreSQL 16 OLTP + 52M+ award rows in ClickHouse OLAP

Filament

2 projects

Primsell

1 project

LITSLINK

1 project

GeeksForLess Inc

1 project

TACANS

1 project

Other Projects

1 project