SWARMCORE
A generative portrait of a multi-agent AI mind — a human face made of live code.
SWARMCORE is generative data-art built on a real system. A real multi-agent pipeline ran — an orchestrator model split a task into three roles (research, analyze, write), three sub-agents executed it with live tools (web search + a knowledge base), and a human steered one tool call mid-flight. Every step was captured to a trace, which SWARMCORE replays as a human face rendered entirely from a field of monospace characters: the face resolves out of the noise as the agents work, words from the trace drift through the code, and the human-steering moment flushes the whole face amber.
The face is made of the run
Not a picture with an overlay — a procedural luminance mask maps per-character brightness on a canvas, so the portrait literally materializes from the swarm's activity.
Replayed, not simulated
Orchestrator planning, sub-agent intents, tool_use blocks, and results are captured verbatim from live API calls and replayed in timestamp order.
The amber moment
When the human-steering checkpoint fires, the entire face flushes amber while a side terminal prints the exact edit — the person inside the loop, made visible.
If SWARMCORE is the portrait of agents working, this is the flight recorder.
A transparent middleware proxy for the Model Context Protocol. It sits between any MCP client and any MCP server, forwards every tool call untouched, and records a full audit trail — arguments, response, duration, success/failure — to a local SQLite database. The proxy exposes its own MCP tools, so the agent can query its own call history at runtime.
17 audit tools for agents
Connected agents introspect their own history — per-tool stats with p50/p95 latency, error rates, slowest calls, argument search, JSONL/CSV export, retention purge.
7 observability tools
One proxy, four backends: Prometheus, Tempo/Jaeger, AlertManager. obs_investigate fans out across all of them concurrently and returns a structured incident summary.
Zero-ops by design
SQLite in WAL mode — no external DB, nothing to operate. Synthesized /metrics, /healthz for k8s probes, bearer auth, Docker image, full CI.
stonks PRIVATE · LIVE
An always-on ecosystem of scheduled AI agents autonomously trading a small, strictly-bounded personal account: an idea-scout hunts candidates across five lanes each morning, an hourly engine manages the book against broker-resident stops, open/close managers bookend the session, and a weekly review grades the process — and evolves its own playbooks inside hard guardrails. Structurally risk-capped: cash account, long-only, no margin, so the worst case is mathematically bounded.
resume-copilot PRIVATE
Chrome extension that reads a job description from the active tab, calls the Claude API with structured tool-use output, and generates a tailored, ATS-optimized resume PDF client-side — including keyword gap analysis against a master profile.
more on github
Earlier work in Java and Python, plus a full-stack database course project from ASU. The recent work — agents, tracing, generative systems — lives in the repos above.