Projects

The systems behind the delivery.

Two projects I designed and built end to end, then the rest of the work underneath them.

Arcus Code

Public

Hardened distribution of opencode, the open-source TypeScript coding agent

Built the model-routing layer upstream lacks. Opencode resolves a model you already named; mine decides per request on cost, latency and capability, so tuning a deployment is configuration rather than a code change.
Added vLLM as a first-class provider and the retrieval that makes it worth using: the agent runs on self-hosted open-weight models inside the customer's network, grounded in RAG over their own corpus.
Rebuilt the MCP tool-calling path so the agent's tool surface comes from the servers attached to a deployment rather than from what shipped in the binary.
Built the evaluation harness that decides whether a change improved the agent, with objective per-case verification, no model grading its own work.
Turns captured agent transcripts into training data with supervised fine-tuning and reinforcement-learning paths on top.

Public

14-service AI career platform, Python and TypeScript

Designed the platform architecture as a reusable substrate, then specialized it into a vertical product without a rewrite.
Agentic workflows behind a four-gate pipeline: permission check on every tool call, verified completion against the tool ledger, deterministic grounding against the corpus, fail-closed guardian.
Multi-agent task division: splits a job across agents working in parallel rather than driving one agent through it serially.
Full-stack: front-end React and TypeScript, FastAPI and Python behind it, PostgreSQL and pgvector underneath serving RAG, every service containerized with Docker.
Cloud architecture on AWS: a provider-agnostic model proxy fronting Bedrock, and a GitHub MCP server held read-only with the token outside the agent.
750+ test functions, constant-time key comparison on every service boundary, append-only audit schema.

Other work

openagent-code
Public, Apache-2.0
A self-hosted autonomous coding agent with a plan, act, verify loop, subagents, cross-session memory, and a permission engine with fail-closed approval review.
View the repository
OpenAgent
Public, Apache-2.0
A multi-service agent platform: identity gateway, LLM-powered APIs, RAG memory on pgvector, and HMAC-signed append-only capture enforced by database role privileges.
View the repository
BoeNet
Private, available on request
From-scratch conditional-compute research in PyTorch, implementing Mixture-of-Depths token routing with top-1 Mixture-of-Experts and Switch load balancing.
DatasetForge
Public, Apache-2.0
A validation-first pipeline that streams and samples Hugging Face corpora into resumable, checksummed pretraining mixes.
View the repository