Invoice Agent

Agentic invoice extraction with per-supplier specialization. Wireframes for the TMLS hackathon demo.

6-beat demo arc

Run the demo arc

Click any beat to open that screen. Each screen has a strip at the top with the presenter script and Prev / Next links so you can rehearse end to end. Total runtime: 5–7 minutes.

1

Queue

/invoice-agent/queue

β€œWe process invoices at scale. Thousands per supplier across hundreds of suppliers. The long tail is where every other tool fails.”

2

Review

/invoice-agent/review

β€œOpen one invoice. Bounding boxes show what came from where. Per-field confidence. Inline edit. Verify checkmark. This is one human pass, not a re-extraction.”

3

Schema editor (add field live)

/invoice-agent/schema

β€œJudge asks "can you also extract po_number?" We add one row in the schema. The next extraction picks it up. Zero retraining. Try it on Nanonets.”

4

Registry (the moat at config-time)

/invoice-agent/registry

β€œEvery supplier has its own mapping. Different fields, different prompt, different model. The supplier-id agent classifies first, the registry resolves the config, the extractor runs against that exact config.”

5

Memory (the moat at runtime)

/invoice-agent/memory

β€œAs humans approve, each supplier moves cold β†’ warming β†’ specialized β†’ locked-in. Accuracy rises, cost drops. Per supplier, automatically. Not a model retrain β€” an orchestration outcome.”

6

Compare (the wedge)

/invoice-agent/compare

β€œSame invoice through Mistral OCR, Qwen-VL, Pixtral, and our LayoutLMv3 specialist. We pick the right tool per supplier. For FGF’s volume, the cost difference is six figures a year.”

✦

Honest tradeoff (narrator)

β€œWe do not auto-pay anything. Humans approve every JSON object before it hits the ERP. Per-supplier specialization assumes recurring volume β€” for a 10-employee company with 50 one-off suppliers a year, our wedge disappears. We are not promising magic on the first invoice. We promise the second one nails what the first one fumbled.”

β–Ά

Roadmap close (narrator)

β€œERP plug-ins next, then anomaly detection β€” this supplier raised prices 12 percent silently, flag it β€” then cross-supplier procurement intelligence. We built a system, not a demo.”

Other screens

Reachable from the side menu, but not part of the rehearsed demo arc.

Build status

Wireframes

7

all screens live on flows.ideaplaces.com

Backend endpoints

9

live at invoice-agent.ideaplaces.com/api/*

Deploy mode

CI/CD

push to main = auto-deploy ~3 min