DK ← BACK TO SITE
INDEX — 07 PROJECTS

EVERYTHING
I'VE BUILT

Search engines, video models, financial agents and betting pipelines. Shipped between classes, hackathons and 3am. Every one of them has source you can read.

01 AI / SYSTEMS / DISTRIBUTED

NEURAMESH PROTOCOL

An Apple-native peer-to-peer protocol that meshes Mac, iPhone and iPad over the local network and shards a transformer's layers across them — so several devices together run a model none of them could hold alone. Hand-rolled UDP transport, Noise-IK handshake and AES-256-GCM encryption, with zero third-party dependencies.

  • Distributed inference — Llama-2-7B sharded across a 3-device mesh
  • 2.6× faster than TCP+TLS on encrypted inference traffic
  • Quorum-sealed distributed memory with K-of-N erasure coding · 474 tests passing
SWIFT RAW UDP NOISE-IK AES-256-GCM LLAMA.CPP
GITHUB ↗
TERMINAL / MESH
02 AI / PRIVACY / SEARCH

CONTEXTUAL

A privacy-first semantic search engine built to replace Spotlight. A local LLM parses each natural-language query on-device — search reaches every local file and nothing leaves the machine.

  • Natural language search across all local files
  • Real-time indexing with minimal performance cost
  • macOS native integration with a system-wide hotkey
PYTHON SWIFTUI OLLAMA QWEN2.5-CODER UNIX SOCKETS
GITHUB ↗
NO CAPTURE YET
03 AI VIDEO EDITING / SAM2 / WHISPER

CLIPPI

3RD PLACE — AGENTIC AI INNOVATION CHALLENGE 2026

A video engine you edit by talking to it. Mistral parses the request, SAM2 segments the frame, and the cut comes back — describe the clip you want and it assembles it.

  • Natural language editing commands
  • Automatic scene detection and object tracking
  • Transcription-based timeline search
MISTRAL AI SAM2 FFMPEG WHISPER FASTAPI CUDA
GITHUB ↗ DEMO ↗
Clippi interface Clippi timeline
04 LLM / FINANCE / CONTEXT-AWARE

PIGEON

BUILT AT TARTANHACKS 2026 · CMU

A financial advisor that knows where you are and how you actually spend. Behavioral patterns plus local cost-of-living produce advice that arrives before the mistake, not after it.

  • Behavioral pattern recognition on spending
  • Budget adjusts to local prices in real time
  • Proactive alerts on unusual activity
REACT NATIVE FASTAPI PLAID API XGBOOST POSTGRESQL
GITHUB ↗
Pigeon dashboard Pigeon insights
05 QUANT / ML / BACKTESTING

EZPARLAYS

Quantitative NFL prediction with a parlay optimiser on top. Gradient-boosted models over player stats, weather and live odds, backtested before anything gets recommended.

  • Backtest hit-rate — ~65% across high-volume bet pools, ~85% on selective high-confidence picks
  • Parlay optimisation across combinatorial space
  • Live odds tracking and arbitrage detection
  • Expected value calculated for every bet
XGBOOST PANDAS SCIKIT-LEARN AIRFLOW FASTAPI
COMING SOON ↗
EZparlays board EZparlays model output
06 ANALYTICS / APIS / DATA

URBAN VITALS

WINNER — BEST STUDENT HACK · STARHACKS 2025

A dozen-plus public data sources — crime, transit, air quality, weather — geocoded and normalised into one livability score per neighbourhood, weighted by what you actually care about.

  • Scoring across 13+ sustainability indicators
  • Interactive neighbourhood comparison
  • Historical crime trend visualisation
MAPBOX GL OPENSTREETMAP EPA AIR QUALITY FASTAPI REACT + VITE
GITHUB ↗ DEMO ↗
Urban Vitals map Urban Vitals scores
07 CLI / GEMINI / SMART CONTRACTS

MIDNIGHT CLI

A command-line agent that uses Gemini to review, debug and auto-fix smart contracts for the Midnight protocol. Point it at a contract, or drop into interactive chat and let it walk the fixes with you.

  • LLM-driven contract review and auto-fix
  • Interactive chat mode for guided debugging
  • Ships as a global npm command over a Python core
PYTHON GOOGLE GEMINI NODE.JS MIDNIGHT
GITHUB ↗
TERMINAL ONLY
THAT'S ALL OF IT

WANT THE NEXT ONE
BUILT WITH YOU?

EMAIL ME ← BACK TO SITE