ArchLens AI

Problem worth solving

Enterprise modernization decks are full of recommendations nobody can falsify — prose findings linked to prose excerpts, with no check that a source exists or that it supports the claim. ArchLens starts from a different contract: evidence before recommendation. If the verifier cannot prove a claim from committed fixtures, the UI refuses to advise — not just flags low confidence.

Intended audience

Architects, engineers, and hiring audiences evaluating how you would ship privacy-first, agentic architecture intelligence. The public demo uses synthetic Northstar Retail fixtures only — no customer or employer data required.

Constraints that shaped the solution

• Public demo must run without API keys or confidential uploads
• Findings cannot live as a second hand-authored copy beside verified output — that would reintroduce unfalsifiable claims
• verified means verifiable right now — stale if a span fails at render time
• Refused claims strike through the recommendation and show No action is proposed from an unproven claim

Architecture decision

Clicking Analyze sample architecture calls a server action that re-reads and re-hashes fixtures/northstar/, runs claims through a verifier registry (first verifier: shared SQL write-target scan), and returns adjudicated findings with evidence spans. The client simulates a multi-agent workflow and human approval gate — but the evidentiary core is deterministic and test-covered (22 vitest tests, GitHub Actions CI).

Proof it is not canned

Edit a fixture file in the repo, redeploy or run locally, and click analyze again — output changes. That is the answer when someone asks whether the demo is replaying static copy.

What was delivered (slice 1b)

• Committed SQL fixtures + sha256-indexed evidence spans
• Verifier outcomes: verified · unsupported · stale
• Planted false claim visibly refused in the UI
• Roadmap approval gate + audit trail (synthetic demo)
• Live at archlens.weidong-shi.com — AI in Action App #5

What comes next

Slice 2: a model proposes Claim objects only — every claim still passes the same verifier before the UI shows it. Hybrid RAG and durable multi-agent orchestration follow once the gate story is solid.

Sibling case studies: Readiness #3 · HabitCheck #4.