Systems engineer building runtimes, integrations, and tools for complex, stateful systems.
I build around external processes and imperfect boundaries: durable state, reconciliation, cancellation and recovery, protocol design, evidence capture, and operator-facing controls. I work primarily in Rust, Python, and TypeScript and have owned systems from architecture and integration through production debugging and operations.
My recent work includes applying that systems discipline to external AI agents and agent-facing protocols. The focus is not another model wrapper; it is the operational infrastructure around long-running work.
Rust-first infrastructure for systematic-trading research, replay, simulation, runtime decisions, and evaluation.
- unifies accounting behavior in an event-driven, one-account state machine with explicit invariants;
- turns large predicate-search spaces into staged bitset and graph search, reserving canonical simulation for finalist verification;
- reconciles desired exposure against observed broker state and retains input lineage, evaluation contracts, and reproducible artifacts.
A Rust research prototype for supervising external AI agents as long-running, unreliable processes.
- models durable attempts, interruption, generation fencing, and SQLite-backed recovery;
- records effects before dispatch, preserves receipts, and represents ambiguous outcomes explicitly;
- distinguishes hosted, managed, and external authority while keeping observed workspace changes separate from agent claims.
It is an experiment in lifecycle, authority, and evidence—not a production security boundary or a universal agent framework.
A Rust-native MCP gateway that supervises stdio servers behind bounded Streamable HTTP, propagates cancellation, and owns child-process cleanup. It validates routes, origins, schemas, and structured results while recording capability routing and execution decisions in SQLite.
A self-contained Rust MCP server that exposes live market data, recording, and authenticated trading through explicit capability profiles. Mutation requires separate opt-in and per-operation confirmation; financial values and large identifiers remain exact across the JSON boundary.
A synthetic Go environment for request-time policy, behavioral correlation, adversarial replay, analyst operations, and intervention review. It keeps false positives, missed abuse, review capacity, and detector limits visible rather than presenting synthetic results as production accuracy.
A backendless WebGPU market-depth instrument driven by public exchange data. It reconstructs the order book from REST snapshots and sequenced WebSocket diffs, detects gaps, retains bounded history in IndexedDB, and renders it in one GPU pass.
I have built and operated SaaS products, API integrations, browser-automation infrastructure, data pipelines, internal tools, dashboards, and financial systems. One internal futures workstation coordinated more than ten accounts from a single operator surface with per-account state, risk controls, health checks, kill switches, replay workflows, and operational review.
Across those domains, the recurring work has been similar:
- turn ambiguous operational requirements into explicit interfaces and state;
- integrate external systems with different failure and authentication models;
- make retries, partial failure, cancellation, and recovery visible;
- shorten the loop between an incident, an explanation, and a verified fix;
- preserve enough evidence for another engineer to inspect what happened.
- Languages: Rust, Python, TypeScript, Go, Ruby, SQL
- Systems: async services, CLIs, WebSockets, JSON-RPC, MCP, REST, SQLite, PostgreSQL, ClickHouse
- Infrastructure: Linux, Docker, Kubernetes, AWS, CI/CD, Grafana, Prometheus
- Product surfaces: developer tools, dashboards, browser extensions, operational control panels