02 — Projects
Tract
Personal Project · Feb 2026 - Mar 2026
- Built Tract, a from-scratch git-style version-control engine for LLM context in Python — a content-addressed commit DAG with branching, three-way merge, and rebase
- Drove the engine from an async agent loop exposing 28 LLM tools, with token-budget enforcement and auto-compression, on the Anthropic SDK and an OpenAI-compatible client
- Added LLM-mediated semantic merges and context compression, mapping model output to typed Pydantic results with fail-open validation
- Persisted the object store in SQLite (SQLAlchemy, auto-migrating schema), backed by a property-based and end-to-end pytest suite
Python · Anthropic · Pydantic · SQLAlchemy · SQLite · pytest
Swarm RAG
Research Project · Dec 2025 - Feb 2026
- Built a GraphRAG retriever (PyTorch) in which an ant-colony swarm of agents traverses the 129K-node STaRK Prime knowledge graph, composing per-step movement, pheromone-deposit, and ranking heuristics to converge on the subgraph most relevant to each query
- Made the swarm's heuristics themselves the search target: each genome encodes its scoring rules as either a weighted sum of primitive graph signals or a free-form expression tree discovered by symbolic regression, so the optimizer discovers traversal logic rather than hand-tuning weights
- Drove the outer search with MAP-Elites quality-diversity — an archive of elite genomes spread across behavioral descriptors (complexity, recall, latency) — explored by a registry of composable genetic operators (tournament/Boltzmann selection, subtree crossover, focused & guided mutation) and a three-tier LLM loop that diagnoses genomes and prescribes targeted mutations
- Implemented it GPU-native in PyTorch, stepping the entire swarm in parallel as batched tensor ops over a compact CSR graph (~16M edges) at ~50 ms/query, with fitness caching, a cross-generation embedding cache, and convergence detection keeping multi-hundred-generation runs fast and VRAM-bounded
- Designed a pluggable architecture with swappable vector-store, graph-store, and embedding-provider adapters (Cohere, Gemini, OpenAI)
PyTorch · NumPy · OpenAI · Gemini · Cohere
KAN-CPPN
Personal Project · Feb 2026
- Implemented from scratch a vectorized B-spline Kolmogorov-Arnold Network layer in PyTorch (spline degrees 1-4, autograd-safe), verified against SciPy as a partition of unity
- Used the KAN as the substrate for a CPPN that synthesizes images from pixel coordinates, reproducing a target genome to 9e-4 MSE over a 2,000-step run
- Built a unified harness comparing three optimizers on the same network: SGD, particle-swarm optimization over spline coefficients, and a natural-evolution-strategy loop with antithetic sampling plus SGD refinement
- Hypothesized that learnable KAN spline activations would be more interpretable than an MLP-based CPPN, built curve-fitting and weight-reset probes to test it, and measured no clear interpretability gain
PyTorch · NumPy · SciPy · pytest
VolleyClip
Personal Project · Jan 2026 - Feb 2026
- Built VolleyClip, a full-stack app (Next.js + TypeScript + FastAPI) that turns a volleyball VOD into per-play clips of a chosen player from a natural-language query
- Used the Twelve Labs video-AI API (Marengo, Pegasus) with schema-constrained prompts to extract exact start/end timestamps for the queried plays
- Engineered an FFmpeg pipeline that stream-copies clips straight from Cloudflare R2 presigned URLs for near-instant lossless cuts, auto-generating thumbnails
- Deployed the frontend on Vercel and the Docker-containerized backend on Google Cloud Run, backed by Supabase Postgres, auth, and row-level security
Next.js · TypeScript · Tailwind · Python · FastAPI · Twelve Labs · FFmpeg · Cloudflare R2 · Supabase · Docker · Vercel
Stitch
UofTHacks 13 · Jan 2026
- Built an agentic AI video editor (Next.js, TypeScript, Supabase) where a Gemini agent edits a live timeline from natural language through 14 custom video and audio tools
- Integrated Twelve Labs semantic search so the agent locates the exact clip segment matching a natural-language query and drops it on the timeline
- Created a Veo 3.1 transition tool that frame-interpolates between the boundary frames of two adjacent clips for a seamless cut
- Generated AI voiceover with ElevenLabs TTS, fitting each clip to a target duration via FFmpeg time-stretching
- Built a custom React 19 timeline editor where AI and manual edits share a single undo/redo history
Next.js · React · TypeScript · Gemini · Twelve Labs · ElevenLabs · Supabase
RLM++
Personal Project · Jan 2026
- Built a long-context QA agent (Python, Agno) after MIT's Recursive Language Models, indexing documents into a hierarchical SQLite summary tree and offloading raw-text reads to sub-agents to stay within the context window
- Engineered a thread-safe round-robin rotation across 14 free-tier LLM endpoints (Gemini, Cerebras, Groq, OpenRouter) with automatic failover, parallelizing summarization of multi-million-token documents under rate limits
- Added a self-healing validator that detects corrupted summaries (provider errors, reasoning-tag leakage, orphaned nodes) and regenerates them in parallel, making ingestion resumable and fault-tolerant
- Benchmarked it with a resumable harness for LongBench-v2 and OOLONG, scoring 78% on an 18-question code-QA slice where the source paper reports base models near 24%
Python · Agno · Gemini · Cerebras · OpenRouter · Anthropic · SQLite
GraphRAG Codebase Agent
Solana Swift SDK · Nov 2025 - Dec 2025
- Built a GraphRAG assistant over the Solana SDK, indexing 600+ Swift symbols into a FalkorDB code graph with typed relationships (inherits, conforms-to, member-of)
- Engineered custom Agno tools for multi-hop graph traversal and vector search with Cohere reranking
- Built a multi-agent Agno implementer that autonomously writes SwiftUI from the repo via GitHub and sandboxed filesystem tools
- Orchestrated cheap open-weight models (GLM-4.6, GPT-OSS-120B) on Cerebras to drive the agents at low cost
Python · FalkorDB · Agno · Cerebras · Cohere · Docker
Solana Swift SDK
Oct 2025 - Nov 2025
- Developed Solana's first official native Swift SDK enabling iOS apps to integrate multi-wallet functionality for leading wallet providers: Phantom, Backpack, and Solflare
- Built the wallet connect and signing flow as an encrypted deeplink handshake with per-session shared secrets between app and wallet
- Architected a protocol-oriented wallet layer so new wallets plug in via conformance, polyfilling Phantom's deprecated sign-and-send over the RPC client
- Shipped the open-source SDK via Swift Package Manager with full documentation
Swift · Solana
SolShare
Hack the North 2025 · Sept 2025
- Developed a bill-splitting iOS app in Swift that automates receipt parsing and payment processing
- Extracted receipt data using Cohere's vision and reasoning models through a self-critic workflow
- Built the backend using Firebase Cloud Functions and Firestore and the frontend using SwiftUI
Swift · Firebase