Rumpy Execution

The public execution layer of my quantitative trading research platform: a cost-aware convex portfolio optimizer (second-order cone program with a market-impact cost model) and a holdings-based backtester, in Rust. Optimality is independently verified rather than trusted: KKT-condition probes against closed-form references, dimensional-analysis checks, and byte-level cross-platform float determinism. 177 tests.

Casus.fyi

AI exam-prep platform that generates board-style practice questions from students' uploaded course material; reached 100+ users through shareable question sets. Schema-constrained LLM pipeline with streaming and embedding-based dedup, passwordless magic-link auth with fail-closed authorization, Redis rate limiting, Sentry/PostHog observability, and 7-workflow CI/CD with staged deploys.

Iris

Self-hosted AI assistant that runs my calendar, tasks, and daily planning over Telegram. A TypeScript MCP tool server with a deterministic planner and a security envelope: prompt-injection quarantine, approval-gated writes, git-backed memory. A model-eval harness replaying its own captured traffic drove a default-model switch that cut per-turn cost by roughly two-thirds. 294 tests.

Maestro

Production runtime for a DAG of trading agents (NestJS): compile-time-typed graph wiring where invalid connections fail the build, staleness gates and circuit breakers, idempotent order submission, and an append-only event log that feeds the research loop. Hexagonal architecture with CI-enforced boundaries; validated on testnet. 994 tests.

Rux-ML

Config-driven ML workbench for gradient-boosted models (XGBoost, LightGBM, CatBoost): Optuna hyperparameter search with purged, time-aware cross-validation for leakage control, full per-trial reproducibility provenance (config hashes, data hashes, seeds, environment), and a gated model registry with atomic champion promotion. 407 tests.

PP-Editor

AI-driven PowerPoint editor built on a hard constraint: anything the AI doesn't touch passes through untouched. Decks live as their raw OOXML tree with a JSON editing view projected on top; edits splice back with hash-verified passthrough regions and Microsoft's OpenXML SDK as the validation gate. The repo includes an end-to-end tour with before/after renders.

Vid-To-Text

Turns video into a timestamped speech/visual/sound timeline using local models (Whisper on CPU, Qwen3-VL via Ollama on GPU), with ffmpeg chunking, transcript-aware vision prompting, and crash-resumable jobs. A Rust client/server pair; nothing leaves my machines except an optional formatting step.

Hospital In A Box

Simulates a very common hospital integration task: ingest HL7 v2 messages (e.g., ADT admit/discharge/transfer and ORU lab results), transform them into FHIR resources (Patient, Encounter, Observation), store them, and show a simple web UI timeline (“Patient admitted → lab results posted → discharged”).

ZerePy

ZerePy is an open-source framework for AI agents written in Python. I collaborated with several other talented engineers to build this framework during my time at Blorm. I was responsible for platform integrations as well as refactoring the codebase to be more efficient and enforce strict typing to reduce bugs.

Blormmy

Blormmy is an AI-powered chat assistant that performs actions on behalf of users. I led backend development, building the service from scratch on a hexagonal/vertical-slice architecture (NestJS/PostgreSQL): per-user agent routing, persistent action history, secure APIs, automated deployment of agent instances to isolated compute, and a plugin layer that processes agent commands from platforms like X and Discord.

Paystand

Involved with numerous projects across ~2 years at Paystand. I designed a zero-fee, instant-transaction payment system for large cross-border B2B payments (hexagonal architecture, plug-and-play providers), working directly with the VP of Engineering on architecture, CI/CD, and launch. I also led a small team building secure multi-approval enterprise funding flows, and built and operated a service anchoring transaction state for auditability. The lessons and skills I learned there have been priceless.

Rarity Ranker

An NLP + Elo-rating system built during my time at Treasure, which generated tens of millions of ranked game items with tuned rarity distributions. Written in Rust with heavy testing.

Club Cards

An expandable digital card collection. I wrote the back-end: the protocol contract and the API that handles claims. Transaction costs were a major constraint during development, so efficiency was the priority: claim authorization is handled off-chain and signatures are verified on-chain, with bit-level optimization to cut storage and transaction costs.