Things I've Built
A collection of products I've shipped — from App Store apps to AI-powered web platforms and internal tools.
MiniLSM
Crash-safe LSM-tree key-value storage engine
A crash-safe LSM-tree storage engine built from scratch in Python, with a checksummed write-ahead log, torn-write recovery, Bloom filters, and streaming compaction.
Built a crash-safe LSM-tree storage engine from scratch with zero lost acknowledged writes across 1,000 randomized SIGKILL trials (~2.2M operations), using a checksummed write-ahead log, torn-write recovery, and atomic file renames. Bloom filters (10 bits/key) cut disk reads per missing-key lookup 99% and raised lookup throughput 4.5×. Streaming k-way merge compaction reduced disk usage 54% and missing-key p50 latency ~40×. A negative test proved the crash harness detects real data loss, and each optimization was checked against saved baselines.
ChefCoach
GPT-4o vision pipeline with rate limiting & CI
An App Store iOS app with a GPT-4o vision pipeline for fridge scanning and recipe generation — backed by token-bucket rate limiting, an LRU cache, JWT auth, and GitHub Actions CI so API cost stays controlled and regressions get caught before release.
Shipped ChefCoach with a GPT-4o vision pipeline and a local fallback for API outages. Built a GitHub Actions CI pipeline with Vitest covering scan quotas, allergy filters, and malformed AI JSON. Engineered a token-bucket rate limiter and an LRU cache (HashMap + doubly linked list) on the recipe-generation endpoint to cut redundant OpenAI calls and cap per-user API cost, alongside JWT auth and input validation. Stack: React, TypeScript, Capacitor, Supabase, OpenAI, RevenueCat.
Applyfy
GPT-4o job platform with AWS resume-parsing backend
Built Applyfy, an AI job platform integrating OpenAI GPT-4o across 12 features, protected by Redis-backed rate limiting (5 req/min/user), reaching 36 signups.
Engineered a serverless AWS resume-parsing pipeline (API Gateway, Lambda, S3, DynamoDB, WAF, CDK). Diagnosed and isolated two independent throttling layers — WAF edge rate limiting vs. Lambda account concurrency — under 50-request concurrent load, sustaining 45 req/sec at 1.9s latency for legitimate traffic.
Collaborative Whiteboard
Real-time multiplayer canvas with CRDTs & Redis
A real-time collaborative whiteboard where multiple users draw and edit simultaneously with conflict-free sync. Built for concurrent collaboration — CRDTs keep every client consistent without a single source of truth locking strokes, while Redis handles low-latency event fan-out across sessions.
I designed the sync layer around CRDTs so concurrent edits merge cleanly instead of overwriting each other — strokes, shapes, and cursors stay consistent even when users draw at the same time. Redis pub/sub broadcasts board updates to connected clients with low latency, and Redis also backs ephemeral session state for active rooms. The Redis instance runs in Docker so local development and deployment share the same reproducible setup. Load-tested the system with 100+ concurrent users collaborating on a shared canvas without breaking sync integrity.
Satellite Telemetry
Satellite tracking with AWS-backed telemetry storage
A satellite telemetry website that tracks latitude, longitude, and speed for satellites in space, with AWS DynamoDB integrated to store and retrieve telemetry data.
Built a satellite telemetry web app that tracks latitude, longitude, and speed for satellites in space. AWS DynamoDB stores and retrieves telemetry so the cloud data layer can scale with the tracking workload.
Brain · AIN Ventures
LLM-powered startup pitch deck analyzer
An internal LLM system used at AIN Ventures to review startup pitch decks and apply AI-driven filtering to pre-seed investment opportunities. Handles unstructured founder data, extracts investment signals, and generates structured investment memos.
At AIN Ventures, I use Brain — an internal LLM pipeline — to process incoming startup pitch decks. The system extracts key signals about the founding team, market size, traction, and product differentiation, then outputs a structured memo that supports investment decisions. I also designed prompt frameworks to improve output consistency.
RAG Document Q&A
Ask questions about any document using LangChain & Pinecone
A retrieval-augmented generation app that embeds document chunks with OpenAI, stores them in Pinecone, and answers questions only from your files — not from the model’s general knowledge.
Loads documents, splits them into chunks, converts each chunk into OpenAI embeddings, and stores them in Pinecone. When you ask a question, semantic search finds the most relevant chunks and GPT answers from those passages only. Built with Python, LangChain, Pinecone, and OpenAI.
Portfolio Website
This site — built with Next.js + TypeScript
A fully custom portfolio built with Next.js, TypeScript, and Tailwind CSS featuring animated particle systems, interactive typewriter effects, smooth scroll animations, and a warm professional aesthetic.
This portfolio was designed and built from scratch. Features a canvas-based particle animation, typewriter effect, glass-card components, and a warm periwinkle + cream color palette. Responsive across all devices.