Multi-agent operations · NYC Tech Week

Student Signal Copilot

Treat the person closest to a problem as both a source of truth and the final verifier that reality changed.

My role

I co-built the multi-agent product, operator experience, confidence model and signal-to-verification narrative with Rosemary Li and Nuo Chen.

View source

Responsibility: Workflow + confidence UX, team of three
Core stack: Express · TypeScript · Claude · SSE
Recognition: 3rd Place Overall · NYC Tech Week

The operating problem

Facilities teams receive fragmented observations and work orders. Recurring issues can remain buried when reports are incomplete, disconnected or never verified after closure.

Evidence

What exists in the current build.

These numbers describe tested product scope—not production impact.

3rdplace overall

8specialist steps per signal

5confidence labels on every claim

4signal-to-verification stages

Inside the build

Screenshots of the running product.

Captured from the application itself—no mockups, no rendered concepts.

Supervisor orchestration trace listing eight completed agents with routing reasons and timings — **Supervisor trace**One sentence in, eight specialist steps out. The operator can see which agent ran, why it was routed there and how long it took.

Confidence ledger labelling each claim Verified, Likely, Inferred or Missing with its basis — **Confidence ledger**Each claim carries its own label and basis. The student’s observation is the only thing marked Verified; everything the model inferred says so.

Verification loop showing a work order closed after the original reporter confirmed the repair — **Closing the loop**A work order only closes when the original reporter confirms reality changed. “Still happening” reopens it and records the false closure.

System workflow

From input to accountable action.

The interface keeps model work, deterministic checks and human responsibility distinct.

Signal

Capture field truth in one sentence.

A student reports what they can directly observe, with an optional photo, before needing to understand facilities terminology.

Evidence

The student’s own observation is the only claim labeled Verified at intake.

01 / System design

More agents did not mean more autonomy.

The workflow decomposes a complex reasoning task into narrow specialists, but deliberately gives none of them execution power.

A deterministic urgent guarantee raises severe, recurring and unresolved issues even if a model misses the pattern.

02 / Outcome

A working operational workflow, not a hypothetical chatbot.

The team built both an operator dashboard and a streamed, inspectable copilot experience grounded in real asset and public data.

The project earned 3rd Place Overall at The City Hacks The State during NYC Tech Week 2026.

Design principles

Ground conclusions before recommending action.
Label uncertainty instead of smoothing it over.
Keep AI read-only and people accountable.
Ask the original reporter whether the fix worked.

Next case study

LabLens