AI products and enablement systems built for real workflows
Orchestration, guardrails, evals, and adoption.
Public projects
My Little Scrum Team
One architecture for a coordinated AI scrum team: one-sentence request through spec, task breakdown, TDD implementation, review, and iteration. Rebuilt across four runtimes: Claude Code plugin, pi.dev extension, VS Code / Copilot participant, and CLI + MCP. Orchestrator write tools are blocked at the tool-hook level, not by prompt. The Blueprint Engine classifies steps as code, LLM, or agent and picks the cheapest node that can do the job.
SREcret-agents
Claude-powered SRE diagnosis agent with a tool-use loop over traces, metrics, and logs. MCP server exposes query_traces, query_metrics, query_logs, correlate_signals, and diagnose. Pluggable TelemetryStore drivers (Jaeger + Prometheus, ClickHouse, composed multi-backend, mock) with an automated boundary test so OTel semantics stay above the backend line. Built with eval logging from day one, paired with a Go OpenTelemetry wrapper and a chaos-injectable test app to stage real “deploy went bad” scenarios. Built as a proof-of-concept to help government partners see how agnostic composition would let them bring their own observability platform without significant rework.
Paper Banana
Six-agent pipeline for proposal graphics: Classifier → Retriever → Planner → Stylist → Visualizer → Critic. Model-per-role routing keeps cheap models on classification and retrieval, stronger models on planning and styling. The Critic runs on a different model family than the Planner to counter self-evaluation bias. Shipped as an MCP server; adopted by the proposal team so exploration became multiple candidate visuals instead of one slow hand-built diagram.
Internal projects
Ad Hoc Market + ahat
GitHub-based marketplace for AI coding-tool plugins, adopted organization-wide. Engineers discover, share, and install Claude Code plugins through one channel instead of tribal knowledge. Companion CLI (ahat) cross-installs the same content into Cursor and Copilot, so enablement is one distribution system across a fragmented tool stack. Shipped with a manifest spec, contributor guide, CI validation, and installers.
Proposal Team
Multi-agent federal proposal system for VS Code Chat (@proposal-team). Specialized agents cover win themes, solution architecture, SME review, compliance (Section L/M), and past performance. Commands walk a solicitation from analyze → Blue Outline → draft → multi-agent review with debate rounds, or /run for the full pipeline with checkpoints. Built so proposal work stays structured and traceable instead of one giant chat thread