Agentic orchestration, LLM tool-use, ML models & inference, and on-device retrieval — on a foundation of systems programming in Rust, C++, and Go.
RSS held flat across a 3-minute stress loop after balancing an Objective-C retain/release leak upstream — the difference between a stable long-running app and a slow crash.
AI Desktop Agent · on-device retrieval
Retrieval-augmented legal information system
Live video streams feed YOLO for object detection; structured detections pass to an LLM that reasons about the scene. Written in C++, Rust, and Python — optimized for low-latency frame processing, stream to LLM output.
Real-time price feeds, sub-second swap execution in Rust, wallet/whale tracking, and live charting. Trade paths optimized for minimal slippage during high network congestion. Next.js/TypeScript order books.
Migrated critical services on a tier handling 15–20M requests/day, keeping p99 on par with the legacy system. Goroutines and bounded channels for backpressure; cutover with zero downtime via gradual traffic shifting.
Fans out concurrent LLM calls and streams tokens back over WebSockets/SSE, cutting time-to-first-token under 300ms. In-memory embedding cache + incremental re-indexing keep vector search under 50ms p99. Hardened in Rust with backpressure, request batching, and timeout/retry.
the problem, owned end-to-end
the right retrieval, the right model
Rust · C++ · Go
Python · TypeScript integration
observability · backpressure · retries
in production, under load
tokio async, lock-free concurrency, zero-copy
goroutines + bounded channels for backpressure
performance-critical, low-latency frame paths
the ML boundary — model integration
product surface, real-time trading UIs
system of record, hybrid stores
hot-read cache, embedding cache
reproducible builds & serving
orchestration for AI workloads
distributed systems infra
service + product APIs
LLM orchestration & tool-use
runtime for on-device inference
tokio async, lock-free concurrency, zero-copy
goroutines + bounded channels for backpressure
performance-critical, low-latency frame paths
the ML boundary — model integration
product surface, real-time trading UIs
system of record, hybrid stores
hot-read cache, embedding cache
reproducible builds & serving
orchestration for AI workloads
distributed systems infra
service + product APIs
LLM orchestration & tool-use
runtime for on-device inference
drag to explore
WithWebview leaked Objective-C retains on macOS/iOSRetained::into_raw() with scoped Retained via Retained::as_ptr()116.5 → 116.9 MB over a 3-min stress loopBuilding agentic AI/ML systems for frontier labs — LLM orchestration, tool-use & function calling, and multi-step reasoning pipelines, plus inference optimization and the MLOps to ship them. Perf-critical infra in Rust/C++, Python at the ML boundary.
Streaming inference gateway with concurrent LLM fan-out; TTFT under 300ms, vector search under 50ms p99, hardened in Rust with backpressure.
PDF → semantic chunking → hybrid vector+keyword store → reranking → cited answers, with vision-LLM ingestion and an autonomous re-indexing agent.
Video-based LLM parser: live video → YOLO detection → structured detections → LLM scene reasoning, in C++/Rust/Python.
Privacy-first AI desktop agent: local semantic engine, on-device embeddings, HNSW retrieval, macOS/Linux binaries idling under 50MB.
Migrated C++ services to Go on a 15–20M req/day tier, p99 held, zero-downtime cutover via gradual traffic shifting.
20–25M+ monthly volume: real-time feeds, sub-second Rust swap execution, wallet/whale tracking, Next.js order books.
Next.js/React/Go/Python/Node. Cut p95 API latency ~40% via query restructuring, N+1 elimination, and Redis hot-read cache.
Production mobile (Flutter, React Native), full-stack web (Next.js, React, Node.js), and blockchain systems (Solidity, Solana/Anchor).
One engineer. Full stack, down to the systems layer.