Software that records evidence as it runs — so humans and agents can inspect, replay, challenge, and safely act on what happened.
By day, I am a Senior Engineer at AWS, building production generative AI applications. Outside work, I create open-source frameworks and research a question I have followed throughout my career:
How can a complex system make its internal state and decisions legible by construction?
My current answer is one design rule:
Record the path at the moment it happens. Never reconstruct it afterward.
Explore the ecosystem · Product view · Technical view
The footprintjs ecosystem applies the same recording rule to four different journeys.
| Journey | System | What becomes legible |
|---|---|---|
| Backend execution | FootPrint | Reads, writes, branches, state changes, and the cause of any output |
| Agent reasoning | AgentFootprint | Context injections, model calls, tool decisions, cost, and failure causes |
| Application interaction | HACI Footprint | The capabilities an agent can currently use, through the signed-in user's real application boundary |
| Exploratory analysis | VizFootprint | Every human and agent path, including abandoned branches and statistical obligations |
Six ecosystem packages are published on npm; VizFootprint is the pre-alpha research system extending the same idea to mixed human–agent visual analysis.
FootPrint — execution provenance
Business logic becomes a directed graph whose execution record is created inline as the graph runs.
- Transactional and patch-based state, parallel fork/join, streaming, checkpoints, and time-travel replay
- Variable-first backward slicing: ask why a value exists and walk it back to the reads, writes, and decisions that produced it
- Auto-generated tool descriptions and causal traces an LLM can reason over
npm: footprintjs
AgentFootprint — context provenance
An agent framework that treats context assembly and every model or tool decision as typed, inspectable evidence.
- Two primitives and four compositions for building agent workflows
- Typed context injections across system, message, and tool slots
- Context-bug localization: shortlist the context that changed an answer, remove it, replay, and confirm the cause
- Causal memory, multi-provider support, MCP integration, pause/resume, and a scored “why this tool?” view
npm: agentfootprint · Live playground
HACI Footprint — human and agent computer interaction
Turns a web application's interaction surface into a typed, traversable skill graph an agent can plan over and use through MCP.
- Starts read-only in guide mode: the agent observes before it acts
- Preserves permissions, preconditions, confirmation gates, and the signed-in user's boundary
- Makes an existing application agent-operable incrementally rather than replacing its UI
npm: hcifootprint · Watch the 37-second demo
VizFootprint — exploration provenance · pre-alpha
Records exploratory visual analysis as an append-only, parent-linked trail shared by humans and agents.
- Every interaction becomes a cause-tagged commit; moving back and acting creates a branch without erasing the old future
- Named paths can be compared, replayed, adopted, archived, and restored
- Statistical memory remains attached to the complete analysis history, including abandoned paths
Read: The Trail Pattern
| Lens | Purpose |
|---|---|
| Explainable UI | Flowchart traversal, time travel, and causal rewind for FootPrint |
| AgentFootprint Lens | Messages, prompts, tools, decisions, and cost on one execution cursor |
| Thinking UI | Scrubbable replay of an agent's reasoning, evidence, and tool alternatives |
My research thesis is that reliable agent transparency should come from recorded decision evidence owned by the framework, not from asking a model to narrate or judge its own reasoning.
- Chain-of-thought: another model-generated claim about what happened
- LLM-as-judge: a second model introduces another trust boundary
- Recorded decision evidence: the runtime owns the trace; humans, smaller models, debuggers, and training pipelines consume the same evidence
Visible Reasoning: User-Facing Decision Transparency for Generative AI Systems — HCI International 2026, LNCS 16745, pp. 3–21, Springer · DOI
Earlier work: Bridging UI Design and Chatbot Interactions — HCI International 2025.
- Enterprise Gen AI Application — practical patterns for production generative AI
- The Flowchart Pattern — when the drawing becomes the program
- The Trail Pattern — when the record becomes the map
- Visible Reasoning — deterministic evidence for human-facing agent transparency
| Stage | What changed |
|---|---|
| HCI and visual analytics | Through my PhD work and Weave, I focused on making complex data-visualization sessions understandable to people. |
| State architecture | StateTree moved the question underneath the interface: if state changes drive the experience, those transitions should be inspectable, comparable, and reversible. |
| Production engineering | Building large software systems made the same problem operational: logs describe fragments, but they rarely preserve the causal structure needed to explain a result. |
| Execution provenance · 2025 | FootPrint made backend code the graph itself, so execution could record its own reads, writes, branches, and causes. |
| Agent context provenance · 2026 | AgentFootprint and Visible Reasoning extended that structure to context assembly, model calls, tool choices, and counterfactual replay. |
| Human–agent application interaction · 2026 | HACI Footprint moved from explaining an agent to defining what it can safely see and do inside a real application. |
| Exploration provenance · 2026 | VizFootprint broadened the work from planned execution to open-ended human and agent exploration, where the record becomes the map. |
| Now | I am focused on provenance-native human–agent systems: software that creates trustworthy evidence during execution and interaction, then exposes the right lens for the person or agent examining it. |
The projects changed. The underlying question did not:
How can software make what happened — and why — available as trustworthy evidence?
PhD in Computer Science, UMass Lowell · Senior Engineer at AWS · Dallas, TX