Areca 360
A ₹250 Cr+ steel service centre runs on Claude. Not a demo, not a pilot — the operating system of a working business at scale. One Claude layer reads the documents, answers the team's questions, and will soon price the credit. It has run every day, unattended, in production.
- Status
- Live · single-tenant · daily use
- Stack
- React · Supabase · FastAPI · Claude
- Claude surfaces
- Parse · MCP · Chat · Extraction · Claude Code
- Record
- Verified — every figure traceable
An ERP records the business — the gap between it and reality leaks into the P&L
A service centre already runs on an ERP. The problem isn't the absence of a system — it's everything the system can't do for itself. The ERP holds the ledger, but it doesn't read the supplier invoice, reconcile the mill certificate, re-check the credit note, or judge the counterparty, and the gap between what the ERP records and what actually moved is where the money quietly goes:
A misread tonne or rate is keyed straight into the ERP. Quantity in TO entered as pieces, a rate off by a digit — and it's now in your GST return, your stock valuation, and the customer's bill. The error surfaces weeks later, in a dispute.
One dropped line on a 100-line credit note is under-claimed input tax. Nobody re-adds a hundred rows by hand before posting. So they don't — and the ITC is left on the table.
Reconciliation is manual, so the ERP lags reality. Books catch up to operations days behind, and by then DSO has crept and working capital is stuck in coil nobody flagged as ageing.
Credit goes out on reputation and a prayer. The ERP shows what a customer owes, not whether they'll pay. On 2–4% net margins, a single bad debtor erases months of contribution — and the risk was sitting in documents the ERP never read.
None of this is a people problem, and it's not a missing-software problem — it's the manual layer around the ERP — the reading, the reconciling, the judging — that exceeds what keying by hand can do accurately. Errors there don't stay operational; they land in tax, in margin, and in bad debt.
What changes when Claude runs it
Document to data, zero keying. Every invoice, CN/DN and MTC becomes structured, sourced line items — each figure traceable to the page and box it was read from. The manual layer around the ERP becomes a data asset.
The arithmetic checks itself. Qty × rate = taxable value is re-derived on every extraction. A mismatch is caught and typed — misread field, dropped line, missing document — before it reaches the books.
Reconciliation is continuous, not a scramble. Operations and the ledger stay in step daily, so ageing and DSO are visible while they can still be acted on.
Risk is priced before credit ships. The layer that reads the counterparty's documents will underwrite the credit they ask for — a traceable verdict in place of a prayer.
Claude is not a feature bolted on — it is the build tool, the intelligence layer, and the agent surface into the business. Five distinct, engineered uses:
Claude as the builder
The entire platform — ERP and credit engine — is engineered, debugged and extended with Claude Code. The loop that builds the product and the loop that operates it are both Claude.
Claude in the runtime
Haiku 4.5 and Sonnet 4.6 parse every supplier document in production — invoices, credit/debit notes, mill test certificates — on a schedule, unattended.
Two-tier model routing
Haiku pre-classifies cheap and fast; extraction escalates to Sonnet only for the hard documents — credit notes past 100 line items. Cost matched to difficulty, not “biggest model for everything.”
Claude behind guardrails
Every extraction is re-derived against a hard identity — qty × rate = taxable value — and errors are typed: misread field, dropped line, missing document. AI plus deterministic verification.
Claude as the team's interface
Anyone on the team asks the live business in plain language — stock, ageing, open orders — instead of digging through screens or pinging the back office. Answered over MCP as the signed-in user, under row-level security.
One intelligence layer, two brains
Document & Operations Brain
Doc brain
Email auto-triage
Gmail pulls invoices, CN/DN and MTCs the moment they land, sorted by type before a human looks.
Parse & self-check
Every PDF becomes sourced line items; qty × rate is re-derived and errors are typed, not just flagged.
Agent-native
The live business is queryable in natural language over MCP.
Credit & Risk Brain · CreditSense
Credit brain
Reads the counterparties flowing through the business and underwrites the credit they ask for — a graded, sourced verdict in place of a guess. Engineered as its own system; the number is deterministic, never an LLM.
Inside CreditSenseCreditSense 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.
The credit brain — prices the risk the paper was hiding
CreditSense reads the counterparties flowing through the business and underwrites the credit they ask for. It is part of Areca 360, engineered as its own system so the engine could be built deep — integration is in progress.
How it helps
It turns “extend credit on reputation” into a graded, defensible call — a letter grade (A–E), a suggested limit, a confidence level, and the green and red flags behind it, each one sourced and timestamped. Something the decision-maker can act on or overrule with full context, never a number to take on faith. It re-runs on a schedule to raise an alert the moment a buyer's grade drops.
Why it's transformative
MSME steel credit is extended blind. CreditSense makes the same call explainable, entity-aware (a partnership isn't docked for lacking a company-only filing), and tunable per tenant — and because Claude does the reading, fast enough to grade every counterparty, not just the worried ones. The operator still makes the credit call; now they make it with the flags in front of them instead of a hunch.
The division of labour — the key design choice
Claude reads; it never scores. Claude Haiku turns messy web and registry results into clean structured findings — adverse media, agency ratings, insolvency mentions, sanctions hits — with name-match filtering, under a strict rule: if it isn't in the source, it returns nothing.
A deterministic engine grades. No LLM in the scoring path — a fixed, published rubric of seven weighted pillars where every grade traces to its inputs. Never a black box.
Two hard safeguards. Fatal signals (dead GST, admitted insolvency, a sanctions hit) force an outright decline — they can't be averaged away. And missing data is labelled “Not available,” never estimated; below a confidence threshold the engine withholds a verdict rather than guess.
The operator decides. Every grade ships with the green and red flags behind it, sourced and timestamped — so the call can be trusted, audited, or overruled with full context. The engine informs the decision; it never takes it out of your hands.
See the full seven-pillar rubric
Identity · resolved first, not scored
The legal and trade name, the entity's registrations and identifiers, its home state and structure, registered address, age and sector — resolved and cross-checked before anything is scored.
Hard stops
The fatal signals — a dead or suspended tax registration, a struck-off entity, admitted insolvency, disqualified or wilful-defaulter directors, an identity that won't reconcile across registries, or personal insolvency of a proprietor or partner. Any single one forces an outright decline.
Compliance
Whether the tax registration is alive and filings arrive on time — return discipline over the trailing year, statutory annual filings, and consistency across every registration held under the same PAN.
Financial
Capital cushion and debt load, the scale of turnover and which way it is trending, and health ratios computed off the filings — leverage, interest coverage and how far payables are stretched.
Legal
Litigation exposure — the active case load, cheque-bounce actions weighted heaviest, insolvency petitions filed by creditors, and formal payment complaints raised against the entity.
Promoter / Ownership
Signals on the people behind the entity — adverse media on promoters, directors stretched across too many companies, and any disqualification or wilful-default history.
Reputation
Market-facing sentiment — ratings and the volume of reviews, recurring themes of not paying vendors, adverse press, and whether the trading address actually checks out.
Stability
How durable the business looks — years in operation, the entity structure, a real web and social presence, a marketplace footprint, and the direction of its workforce.
Payments
How the buyer actually pays — days beyond terms reported by trade references, whether those suppliers would deal with them again, and self-declared or auditor-confirmed dues owed to smaller vendors.
Joins the composite only when real payment evidence exists — absence of data never reads as a negative signal.
Evidence-only
Auditor and going-concern signals, distress in group companies, provident-fund regularity, external credit ratings and cooperation flags, and sanctions screening.
Drive red flags and confidence today; folding them into pillar math is a queued decision.
Every data point carries its source and retrieval timestamp, and scores on a published band — the “why” behind any grade is fully traceable.
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.
What building this taught us
Match the model to the difficulty, not the ego. Haiku classifies everything cheaply; Sonnet is spent only on the documents that actually need it. “Biggest model for everything” is how AI projects die on cost.
AI proposes, arithmetic disposes. Probabilistic extraction is wrapped in deterministic checks. The model's confidence doesn't post to your books — a re-derived qty × rate does.
Type your errors. “Failed” is useless. “Misread field / dropped line / missing document” tells you what to fix and lets the system route the exception.
Null-plus-reason beats a confident guess. A system that says “Not available” is trusted; one that fabricates to look complete is used once.
Pin the engine like flight software. Five synthetic personas form a permanent regression suite — any change that drifts a grade fails the build.
Build and operate loop, same tool. The product is built with Claude Code and run on Claude in production.
Depth in one real business before breadth across many. Built inside a working service centre, by its operator — single-tenant proof is the wedge; the platform is the play.
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.
Screenshots can be staged; arithmetic cannot. Every artefact below is the system checking itself in public.
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.
Edge functions
26deployed
Deployed functions run the pipeline — sync, parse, verify, serve — on schedules, unattended.
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.
Why steel, why India, why now
India's steel flows through thousands of MSME service centres running on ERPs that record the business but can't read the documents or price the risk moving through it. Areca 360 was built inside one of them, by its operator. The document brain removes the paper; the credit brain prices the risk the paper was hiding. One proven single-tenant deployment is the wedge; the platform is the play.
How it's built
- Engineered with
- Claude Code
- Models in production
- Haiku 4.5 · Sonnet 4.6
- Agent surface
- MCP · OAuth · RLS
- ERP stack
- React · Supabase · edge functions · Tally sync
- Credit stack
- Next.js · FastAPI · Postgres · deterministic scorer