By day, Senior Engineer at AWS, building production generative AI applications. Outside work, I create open-source frameworks and write about the hard problems most teams skip when shipping enterprise AI.

Making the internal state of complex systems legible — first to humans, now to AI.

Weave  ─►  StateTree  ─►  FootPrint  ─►  agentfootprint  ─►  hcifootprint
(data vis  (state         (execution    (context            (same
 sessions)  diffing)       graphs +       engineering,        legibility,
                           causal         abstracted;         pointed at the
                           traces)        decision            interaction
                                          evidence            surface — apps
                                          as a cache)         agents can
                                                              operate)

Each step abstracted one more thing the previous step did by hand.

Business logic becomes a directed graph that produces causal traces an LLM can reason over.

• 7 flow patterns · transactional state · PII redaction · auto-generated tool descriptions
• 6 modular libraries: memory · builder · scope · engine · runner · contract
• Parallel fork/join · streaming · patch-based state · time-travel replay
• Variable-first backward slicing — ask "why is score N?" and walk the variable back to its birth · pause/resume checkpoints across servers

npm install footprintjs

The autograd-shaped abstraction for LLM apps. Built on FootPrint. You describe injections; the framework evaluates triggers, composes slots, and observes every decision as a typed event.

• 2 primitives (LLMCall · Agent) + 4 compositions (Sequence · Parallel · Conditional · Loop)
• 1 Injection primitive, 4 typed factories (defineSkill · defineSteering · defineInstruction · defineFact) targeting 3 slots (system · messages · tools) under 4 triggers (always / rule / on-tool-return / llm-activated)
• 1 Memory factory — 4 types × 7 strategies, including Causal (decision-evidence snapshots → cross-run replay, cheap-model triage, training data — three uses of the same recording)
• 47 typed observability events across 13 domains · pause/resume across servers · multi-tenant identity · MCP integration
• $0 dev via mock() · 6 LLM providers (Anthropic · OpenAI · Bedrock · Ollama · Browser variants)
• Context-bug localization — when an agent answers wrong, pinpoint which injected context caused it (walk the trajectory, shortlist early culprits, confirm by replay) · a "why this tool?" scorer over every tool it could have picked
• Live Playground — paste a trace, drag the time-travel slider

npm install agentfootprint

Turns a web app's interaction surface into a typed, traversable skill graph an LLM agent can plan over and act on — served over MCP, so the agent acts as the signed-in user through the app's own handlers.

• Adoption ladder starts read-only ("guide mode") — the agent observes before it acts
• Ships llms.txt for agent consumers · 281 tests
• A vanilla dress-shop becomes agent-operable in three commits (demo repo)

▶ Watch the 37-second demo

npm install hcifootprint

Full graph → footprintjs.github.io

A framework for deterministic LLM-agent transparency. A third paradigm distinct from:

• Chain-of-thought — the agent narrates its own thinking (unverifiable; the narration is more LLM output)
• LLM-as-judge — a second agent grades the first (recursive trust problem)
• Recorded decision evidence — the framework owns the trace. Every decision the flowchart makes is a typed event; humans, cheaper LLMs, and training pipelines all consume the same recording.

Visible Reasoning: User-Facing Decision Transparency for Generative AI Systems — published at HCI International 2026 · LNCS 16745, pp. 3–21 · Springer, Cham · doi:10.1007/978-3-032-30849-8_1. Earlier work — Bridging UI Design and Chatbot Interactions (form-based principles → conversational agents) — published at HCII 2025.

Enterprise Gen AI Application — LinkedIn newsletter, 320+ subscribers · The Trail Pattern — the footprintjs ecosystem blog.

PhD in Computer Science, UMass Lowell · Dallas, TX LinkedIn · Medium ·