[ PILOT]
Built at Quantum Strides · sole engineer

Agentic Walkthrough Assistant for Enterprise IWMS

An agent that autonomously learns a customer's asset-management system by driving its live UI — safely — and turns it into a queryable knowledge graph.

Role
Sole Engineer
Year
2025
Stack
9 technologies
Status
Pilot
// overview
A multi-agent system that onboards itself to a complex enterprise IWMS the way a new analyst would: by clicking through the real interface. Browser automation drives the live UI while Claude reasons over what it sees, mapping pages, forms, and workflows into a Neo4j + Pinecone graph for retrieval — with a hard guarantee it can never mutate customer data.
// the problem
Every customer's IWMS instance is configured differently, with hundreds of forms and undocumented workflows. Manually mapping one took analysts weeks, and any automated crawler risked accidentally submitting or deleting live records. The agent had to learn fast and be provably read-only.
// architecture
01 · drive
Browser agent
Playwright
02 · perceive
Vision + DOM
Claude reasons
03 · extract
Form schema
24–47 fields
04 · graph
Neo4j + Pinecone
GraphRAG
05 · serve
LangGraph chat
walkthroughs
// side rail
Never-mutate guard
code-enforced
⛔ any Submit / Save / Delete click is physically blocked — verified live
// what i built
  • Browser-automation layer (Playwright + browser-use) that drives the live IWMS while Claude interprets each screen.
  • Vision-based form extraction that opens modals and captures fields, types, required flags, tabs, and sub-grids — 24–47 fields per form.
  • A code-enforced never-mutate guard (not prompt-based) that physically blocks any Submit/Save/Delete, verified against production.
  • Single-producer / multi-consumer concurrency to respect the IWMS one-session constraint while parallelizing the LLM step.
  • GraphRAG store in Neo4j (structure) + Pinecone (vectors), and a LangGraph chat pipeline with a guarded live-walkthrough mode.
  • Token-accurate per-run cost instrumentation; shipped under strict TDD with 300+ passing tests.
// impact
2.7×
faster extraction
0+
tests passing
0
mutations, ever
// tech stack
PythonLangGraphClaude (Bedrock)Neo4jPineconePlaywrightFastAPIbrowser-useReact
next project
Multi-Tenant White-Label AI Chatbot