Inspiration
Restaurant operators often discover discount misuse, refund spikes, and labor overruns after the month closes, when the money is already gone. We wanted a daily control loop that surfaces evidence without letting AI invent financial claims or execute operational changes.
What it does
LeakLens Control Room analyzes store-isolated daily metrics, verifies the financial identity, compares current rates with each store's own short historical median, and ranks locations by material scenario impact. It shows the exact evidence and caveats behind labor, discount, and refund exceptions.
Every proposed action starts pending. A human must approve or reject it. Approval executes nothing; it opens a measurement step. SQLite then preserves three distinct audit facts: the original estimate, the human decision, and the measured outcome. The decision locks after measurement.
How we built it
- Deterministic Python owns every rate, threshold, ranking, and dollar calculation.
- GPT-5.6 through the authenticated Codex CLI is an optional bounded narrative provider. It runs with a read-only sandbox, structured output schema, timeout, and no shell execution.
- Provider failures return a deterministic brief with safe provenance instead of leaking an error or changing the numbers.
- A standard-library CLI, HTTP API, responsive dashboard, and SQLite audit store expose the workflow.
- The public Render demo uses deterministic mode, synthetic CSV fixtures, and an ephemeral database.
Challenges
The hardest problem was enforcing a real trust boundary. The model cannot modify findings, dollar values, decisions, or operational systems. We also separated projected scenarios from measured outcomes so the interface never presents an estimate as guaranteed savings.
Accomplishments
- Three-store portfolio isolation and ranking
- Human approval/rejection with no external execution connector
- Append-only measured-outcome workflow
- Provider/model/fallback/report provenance
- 23/23 unit and integration tests passed, including a fresh public GitHub clone
- Public hosted demo and reproducible synthetic dataset
What we learned
For operational AI, a narrow auditable loop is more valuable than an impressive autonomous claim. Deterministic dollars, bounded explanation, and explicit human control make the system inspectable.
Honest scope and what's next
This prototype uses synthetic CSV ingestion and transparent median-plus-threshold rules; it is not seasonal forecasting. The commercial roadmap is Toast/Square POS ingestion, channel-level commission analysis, store-scaled severity, day-of-week and seasonal baselines, and model-aware action prioritization.
Log in or sign up for Devpost to join the conversation.