# SKAI

- **Event:** [RAISE Summit Hackathon](https://cerebralvalley.ai/e/raise-summit-hackathon)
- **When:** Jul 4 at 9:00 AM – Jul 5 at 7:00 PM (GMT+2)
- **Where:** Paris, France
- **Team:** [Amod Adarsh](https://cerebralvalley.ai/u/strell)
- **GitHub:** https://github.com/AMODADARSH/taxagent
- **Demo video:** https://youtu.be/dw-8ox_XEIc
- **Gallery:** https://cerebralvalley.ai/e/raise-summit-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/raise-summit-hackathon/hackathon/gallery/3

Ledger is an enterprise tax compliance agent that audits invoices for indirect tax errors and explains exactly why each one is wrong. Finance teams issue thousands of invoices under rules full of exceptions: different rates by category, special treatment for exports, reverse charge for cross border B2B, thresholds that flip the rules once crossed. Manual review is slow and mistakes are common, and the cost of getting it wrong is real money at audit time.
For every invoice, the agent plans what needs checking based on the transaction's facts, retrieves the governing rule from a Vultr hosted vector store, independently verifies the correct rate using a deterministic calculator tool rather than trusting a language model with the math, cross checks the declared amount against that verified figure, and when there's a mismatch it retrieves again to gather the context needed for a clear explanation. The result is a structured escalation memo: which invoices are correct, which are flagged, which need a human to look closer, and the total tax exposure at stake, each with plain language reasoning grounded in the actual rule text.
The same architecture runs on two completely different tax systems, EU VAT and Indian GST, using the identical agent logic with only the rule corpus swapped out, which shows the design generalizes rather than being hardcoded to one country's tax code.
I also built a side by side comparison mode showing a naive single call RAG answer next to the tool verified answer for the same invoice, so it's visible exactly where the extra planning and verification steps earn their keep, particularly on ambiguous cases where a single call confidently guesses but the multi step agent correctly flags the case for manual review instead.
Built solo end to end, from the rule corpus and calculator logic through the Vultr RAG integration and the frontend.

## More from RAISE Summit Hackathon

- [Taste Engine](https://cerebralvalley.ai/e/raise-summit-hackathon/hackathon/gallery/1)
- [VOID](https://cerebralvalley.ai/e/raise-summit-hackathon/hackathon/gallery/2)
- [Ouday Benabid](https://cerebralvalley.ai/e/raise-summit-hackathon/hackathon/gallery/4)
- [Helius](https://cerebralvalley.ai/e/raise-summit-hackathon/hackathon/gallery/5)
- [CovenantSentinel](https://cerebralvalley.ai/e/raise-summit-hackathon/hackathon/gallery/6)
- [Ryan Nishikawa](https://cerebralvalley.ai/e/raise-summit-hackathon/hackathon/gallery/138)

---

Markdown version of https://cerebralvalley.ai/e/raise-summit-hackathon/hackathon/gallery/3. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt