Areca 360
An AI-native operating system for steel trading & processing — one Claude layer powers its documents, its agents, and its credit decisions.
- Status
- Live · single-tenant
- Stack
- React · Supabase · Claude
- AI surfaces
- Gmail · parse · MCP · credit
- Record
- verified
One intelligence layer, two brains
Document & Operations Brain
Doc brain
Email auto-triage
Gmail fetch pulls invoices, CN/DN and MTCs the moment they land, then sorts them by document type before a human looks.
Document parsing
Every PDF becomes structured, sourced line items — each figure traceable back to the page and box it was read from.
Self-check & classify
Qty × rate is re-derived and errors are typed: misread field, missing line, or missing document — not just flagged.
Agent-native
The live business is queryable in natural language over MCP — ask it about stock, ageing or dispatches directly.
Credit & Risk Brain · CreditSense
Credit brain
Seven-pillar deterministic score
Compliance, financials, legal, promoter, reputation, stability, payment behaviour — a fixed, published rubric where every verdict is traceable to its inputs. Never a black box.
Hard-stop overrides
Cancelled GST, admitted insolvency, a sanctions hit — fatal signals force a decline outright. A deal-breaker can never be averaged out by an otherwise-healthy score.
Confidence-gated honesty
Missing data is labelled “Not available”, never estimated to look complete. Confidence ships beside the grade — below threshold, the engine withholds a verdict rather than guess.
Entity-aware & tunable
A company, LLP, partnership or proprietorship is judged on the rubric that applies to it — and every weight, grade band and limit factor is per-tenant credit policy.
CreditSense is part of Areca 360, engineered separately — the credit-brain integration is in progress, not yet live. Claude powers the reading, not the number — the full engine is detailed in Inside CreditSense.
Seven AI builds inside the ERP
Not a wrapper around one prompt — seven engineered AI systems, each doing a real job inside a running steel service centre.
AI document intelligence
FlagshipUnstructured mill PDFs become clean, structured, reconciled data — no human keying. Supplier emails auto-sync from Gmail; every invoice, credit/debit note and mill test certificate is parsed by Claude into typed rows in the database. The industry's paper problem, turned into a data asset.
Two-tier model routing
A multi-stage pipeline, not a single prompt. Claude Haiku 4.5 pre-classifies every document — cheap and fast. Extraction then routes by type, escalating to Claude Sonnet 4.6 for the hard ones: credit/debit notes running past 100 line items. Two models, matched to cost versus difficulty — not “send everything to the biggest model.”
Domain rules + arithmetic self-verification
The parser encodes steel-document rules — read Quantity-TO tonnes, not Quantity-NO piece counts; check column order; normalise formats — then reconciles Claude's output against a hard identity: qty × rate = taxable value. Extraction errors are caught and classified — misread field, dropped line, or missing document. AI plus deterministic guardrails, so the numbers are trustworthy rather than “whatever the model said.”
Agent-native — the ERP is callable by AI assistants
An OAuth-protected MCP server exposes live business tools — inventory_summary, list_open_orders — so external AI agents (Claude, ChatGPT) query the actual business as the signed-in user, under row-level security. The ERP isn't just AI-assisted; it's an AI-queryable data source, positioned for the agentic shift.
In-app conversational AI
An inventory-chat function brings natural-language Q&A directly onto live inventory data — ask about stock, grades and ageing the way you'd ask a colleague.
Production-grade AI-ops discipline
What separates this from a demo: the AI runs on a schedule, at scale, with the failure modes actually handled — per-document attempt caps against runaway token spend, arithmetic reconciliation for accuracy, model routing for cost, and queue-and-retry design for resilience. AI used as dependable infrastructure.
AI-built and AI-maintained
The application itself is majorly built with Claude Code — engineered, debugged and extended by Claude. The loop that builds the product and the loop that operates it are both AI-driven, end to end.
The credit brain — built separately, integrating in
CreditSense is part of Areca 360 — engineered as its own system so the credit engine could be built deep and rigorous, and now being integrated into the platform. Like the rest of the build, it's developed with Claude Code. The division of labour is deliberate: Claude does the reading, a deterministic engine does the number.
What Claude does
Claude reads
Adverse media & sanctions
News, UN / OFAC sanctions lists and reputation signals — the qualitative desk-work a credit analyst does by hand, done at machine scale.
Entity resolution across 25+ sources
One counterparty resolved into a single, fully-sourced file — every claim traceable to where it was read, so the picture is whole, not scattered.
Never the number
Claude's reading feeds the engine as evidence. The score itself is deterministic — Claude does not generate it, so every verdict is explainable and defensible.
What the engine does
The engine decides
A fixed, published rubric — seven pillars
Payment is gated — it only enters the score when real payment evidence exists, so missing data never masquerades as a negative signal.
Hard stops
- Cancelled GST registration
- Admitted insolvency
- Sanctions-list hit
- Promoter identity mismatch
Any of these forces an automatic decline — a deal-breaker can never be averaged out by an otherwise-healthy score.
Honest, entity-aware, per-tenant
Missing data is labelled “Not available”, never estimated; below a confidence threshold the grade is withheld. A company, LLP, partnership or proprietorship is judged on the rubric that applies to it. And every weight, grade band and limit factor is configuration — each tenant encodes its own credit policy.
Every assessment returns
- Grade
- A – E
- Verdict
- extend · conditions · advance-only · decline
- Limit & terms
- conservative, suggested
- Confidence
- how much is verified
- Evidence
- fully-sourced trail
The same engine re-runs on a schedule for ongoing monitoring — raising alerts when a buyer's grade drops or a new red flag appears.
Evidence, not adjectives
The layer's output is inspectable: a parsed invoice that caught its own error, a transparent credit verdict, and a live business you can just ask.
Parsed invoice
| Line item | Qty × rate | Value |
|---|---|---|
| HR Coil 2.5mm | 12.40×62,500 | ₹7,75,000 |
| CR Sheet 1.2mm | 8.10×71,200 | ₹5,76,720 |
| GP Coil 0.8mm⚠ flagged | 5.00×68,400 | ₹3,08,000 |
| MS Plate 10mm | 3.25×58,900 | ₹1,91,425 |
GP Coil 0.8mm — arithmetic mismatch → flagged & classified (misread field)
Figures illustrative & masked
Credit verdict
Extend credit — with conditions
Payment pillar gated — no verified payment evidence yet, so it stays out of the score.
Seven-pillar deterministic rubric · hard-stop overrides · confidence-gated. Claude does the reading, not the number.
Figures illustrative & masked
Live MCP query
Figures illustrative & masked
Proof it actually runs
Live inventory
5grades
Tracked coil → WIP → dispatch, with full genealogy from the incoming coil to the piece that leaves the gate.
Tally sync
Daily
Books auto-matched against operations every day — the system of record and the ledger stay in step without manual reconciliation.
Dispatches
▪▪▪▪Kg
Every dispatch weighed, logged and tied back to its source coil. Volumes are masked here — the audit trail behind them is real.
Figures illustrative & masked
Run the business, then underwrite the credit that flows through it
Run
Areca 360
The operating system — documents, inventory and agents that run a real steel service centre.
Underwrite
CreditSense
Part of Areca 360, engineered as its own system — it reads the counterparties flowing through the business and underwrites the credit they ask for.
Become
One platform
A single system of record where running the business and pricing its credit are the same motion.
How it's built
- Engineered with
- Claude Code
- Models in production
- Haiku 4.5 · Sonnet 4.6
- Agent surface
- MCP · OAuth · RLS