Projects

Agentic systems, LLM platforms, and the research behind them.

Six systems I designed and built end to end — architecture, data model, evaluation, and deployment.

openagent-code

Public · Apache-2.0

Self-hosted autonomous coding agent.

Roughly 27,000 lines of Python with 25 tools, a plan → act → verify orchestration loop, subagents, cross-session memory, context compaction, and MCP integration. Its permission engine enforces allow/ask/deny rules, a workspace boundary, and a fail-closed LLM approval reviewer. A 23-task evaluation harness with objective per-case verifies keeps it honest, and a LoRA-SFT distillation pipeline promotes a student model only through an automated base-vs-student gate.

  • Python
  • vLLM
  • Amazon Bedrock
  • MCP
View on GitHub

OpenAgent

Public · Apache-2.0

Five-service agentic platform at 135/135 tests passing.

An identity gateway, a provider-agnostic model-routing proxy, LLM-powered APIs, a RAG memory service on pgvector, and HMAC-signed append-only capture with partitioned retention, plus LLMOps observability. Deployed across AWS and self-hosted GPU infrastructure.

  • Python
  • FastAPI
  • PostgreSQL
  • pgvector
  • Docker
  • AWS
View on GitHub

CareerAgent

Private · available on request

A 14-service vertical product specialized from the OpenAgent architecture without rewriting it.

Roughly 37,000 lines, 72 endpoints, five PostgreSQL databases, and about 800 hermetic tests. Its defining feature is that it will not let the model fabricate anything: deterministic anti-fabrication grounding plus a fail-closed LLM verifier.

  • Python
  • FastAPI
  • PostgreSQL
  • Docker

BoeNet

Public · Apache-2.0

From-scratch PyTorch conditional-compute research.

A hand-written transformer, training loop, and sampler implementing Mixture-of-Depths token routing and top-1 Mixture-of-Experts with Switch load balancing, which I call MoDE. It matched the dense baseline’s perplexity at roughly 49% routed-token compute at small scale; getting there meant diagnosing a 1,000× router-gradient suppression and fixing it with a dedicated router learning rate. 46-test suite.

  • PyTorch
  • Python

Arcus Code

Private

My hardened downstream distribution of opencode, an open-source 266,000-line TypeScript coding agent. The distribution itself is private and not open source.

I neutralized a supply-chain hazard where the inherited auto-updater silently replaced the product with the upstream binary, gating self-update, uninstall, and upstream GitHub-App integrations behind a single ownership flag. I root-caused a Bun-on-Windows filesystem bug that broke every plugin install and a lock-lifetime concurrency defect, and shipped a zero-data-loss data-directory migration. Every fix shipped with regression tests.

  • TypeScript
  • Bun
  • React

DatasetForge

Public · Apache-2.0

Validation-first pretraining data pipeline streaming TB-scale Hugging Face corpora.

Produces resumable, checksummed pretraining mixes with validation at every stage, covered by 157 passing tests in CI.

  • Python
  • Docker
  • Hugging Face
View on GitHub

Get in touch

If you’d like to talk about agentic systems, LLM applications, or the infrastructure behind them, email is the fastest way to reach me.