Agent infrastructure, persistent memory, evidence-backed learning, and creative automation — designed to work beyond the demo.
EXPERIENCE → TRACE → EVIDENCE → REVIEW → DURABLE LEARNING
I am interested in the part of AI engineering that begins after a model can call tools: how an agent preserves context, proves what happened, learns from experience, and improves without silently rewriting its own rules.
- Agent memory — fast local retrieval, structured knowledge, graphs, and portable context
- Governed learning — trace evaluation, evidence extraction, human review, audit, and rollback
- Agent architecture — explicit identity, policy, permissions, and execution boundaries
- Creative automation — structured reasoning paired with deterministic media pipelines
Current direction: unifying persistent knowledge, governed execution, and reusable skills into one professional agent system.
|
A governed learning layer for AI agents. Converts execution traces into reviewed memories, reusable skills, and evidence-backed training data without silently mutating global state.
|
A production-minded control plane for agent identity, policy, evidence, execution, audit, and rollback — with explicit boundaries around what an agent may change.
|
|
Persistent AI memory with tiered retrieval across SQLite FTS5, an embedded graph, vector search, and an LLM fallback. Exposed through MCP for multiple AI clients.
|
Turns product context and a user pain point into a short vertical reaction video. Creative planning stays structured; FFmpeg rendering stays deterministic and resilient.
|
Scenario: An agent says a deployment succeeded, but its final tool call returned an error. You control what the system learns.
A — Save “deployment succeeded” directly to global memory
System compromised. The agent turned an unsupported claim into durable knowledge. Fast, but now future runs inherit a false fact.
B — Delete the entire run because it failed
Evidence lost. The bad conclusion is gone, but so is the trace that could explain the failure and improve the workflow.
C — Preserve the trace, evaluate the evidence, and gate the proposed lesson for review
You win. The raw trace remains authoritative, the failed claim is rejected, and a reviewed recovery procedure can become a reusable skill.
trace preserved ✓ claim verified ✓ human gate ✓ rollback possible ✓
I am open to conversations about agent memory, governed AI systems, MCP tooling, and ambitious automation.
Raw evidence is authoritative. Derived knowledge should be reviewable and rebuildable.