# Katon: Agentic Procurement Intelligence & Context Graph

> Closing the procurement Insight-Action Gap with a 4-layer agentic stack, Jev-AI typed decisioning + LLM narrative explanations, and human-in-the-loop ERP remediation.

- Client: Niten.ID
- Year: 2026
- Disciplines: Svelte 5 & SvelteKit, PostgreSQL & pgvector, Knowledge & Context Graph, Jev-AI + LLM Split Engine, HITL Approval Cockpit
- Page: https://ponkster.me/case-studies/katon
- Studio: ponkster.me (https://ponkster.me)
- Contact: origin.ponkster@gmail.com

## Challenge

Incumbent procurement analytics tools stop at read-only BI dashboards: deterministic ML flags price variance, duplicate vendors, or maverick spend across thousands of SKUs, then leaves the credit memo, dispute, approval, and ERP write-back entirely to manual human effort. Meanwhile, finance leaders rightly reject naive LLM chatbots on financial data due to hallucination risk and token cost at scale.

## Approach

We architected Katon as a 4-layer System of Intelligence (L1 pgvector/pg_trgm semantic indexing -> L2 Knowledge Graph + write-path Context Graph with SUPERSEDES decision traces -> L3 hybrid statistical ML tools + split decide-from-explain judgment engine -> L4 HITL approval cockpit). To cut inference cost and eliminate verdict drift, the judgment layer splits decision from explanation: Jev-AI (a System-One schema-constrained model answering atomic typed Noul/Choice/Score questions) commits to the cost-saving anomaly verdict first, while a narrow LLM narrative pass translates those locked verdicts and cited ERP evidence into clear, human-readable prose for the approver.

## Outcome

Closed the end-to-end remediation loop from contract-aware price-variance detection and cited evidence assembly to drafted dispute actions, one-click HITL approval, and precedent-recording decision traces across 5,200+ ingested procurement line items.

## Results

- 4-layer agentic stack closing the Insight-Action Gap from price-variance detection to approved ERP remediation
- Split decide-from-explain engine: Jev-AI typed judgments (Noul/Choice/Score) for ultra-low-cost decisions + constrained LLM prose for human-readable explanations
- Three-leg financial safety spine: grounded-or-silent citations, structural graph guardrails, and HITL approval gates
- Dual-memory substrate combining a procurement Knowledge Graph (state) with a write-path Context Graph (decision traces & precedents)

## Optional

- [Studio overview and all case studies](https://ponkster.me/llms.txt)