- Project
- AI/ML
- Web
- Team
- Oct 2026
Simpul: AI data hub with Gemini agents
A 10-day prototype for a team in an internal corporate AI competition. The team won 1st place.
How it works
- 01Data Steward: Gemini maps columns of any distributor file
- 02Code checks every row, outlet name, and rule; people review exceptions only
- 03Analyst: function calling over 14 query tools, so numbers come from code
- 04Program Auditor: rule-based claim checks, no model in the loop
Query tools for the analyst agent
Every number in an answer is computed by code, not generated.
Columns mapped by Gemini
8 messy distributor files × 3 runs, simulated data; 94.8% of outlet names, zero wrong merges.
Simulated claims matched by the rule auditor
348 of 348.
14 query tools, numbers from code
Pick a question to see which tools the analyst calls. Example routing on simulated data.
No question yet. Every number in an answer comes from one of these tools.
The problem
Distributors send sales files in any shape: spreadsheets, PDFs, chat exports. Analysts lose days cleaning them, and a chatbot that invents numbers cannot be trusted with sales decisions.
My part
Built the React 19/TypeScript prototype, the serverless Gemini endpoints on Vercel, the evaluation runs, and every presentation material in 10 days (21-30 Oct 2026).
I built the prototype and all materials for the team; team members and jury are not named.
What did not work
- Letting the model answer from raw tables invited invented numbers, so every figure moved behind a query tool.
Limitations
- All evaluation used simulated data: 8 files × 3 runs (36 calls, USD 0.147).
- A prototype for a competition, not a production system.
