Shipping an Enterprise LLM Platform for Banks: 7 Lessons from JARVIS RCSA

JARVIS RCSA is an enterprise AI platform that automates Risk & Control Self-Assessment for banks: it ingests policies, SOPs, and contracts, extracts process structures with LLMs, generates process maps and Word/PowerPoint deliverables, and runs verification passes before anything reaches an auditor. I contributed 528 commits across a multi-tenant codebase of roughly 179k lines — the Smart Router orchestration system, process-map rendering, reusable Word/PowerPoint/PDF/BPMN export pipelines, and the verification framework. These are the lessons that transfer to any serious LLM product.

1. Verification is the product

Banks don't buy AI output; they buy defensible output. The verification framework gates every AI result against regulatory standards — SR 11-7 (model risk management), the EU AI Act, DORA, and BCBS 239 — before export. Every claim must trace to a source document. If your enterprise AI product can't answer 'why did it say that?', it will not survive procurement.

2. Route work to models, don't pick one model

The Smart Router sends each task — extraction, summarization, structuring, verification — to the model and prompt configuration that wins for that task type. Model choice is a routing table, not an architecture decision. When a better model ships, you update configuration, not code.

3. Deliverables beat chat

Enterprise users don't want a chat window; they want the Word document, the PowerPoint, the BPMN diagram their process already expects. The export pipelines (docx, pptx, PDF, BPMN) were some of the highest-value code in the platform. Meet the workflow where it is.

4. Queues everywhere

Document ingestion, LLM calls, exports — everything long-running goes through BullMQ workers with retries and per-tenant fairness. Synchronous LLM calls in request handlers is how demos behave; production systems are queues with status.

5. Multi-tenancy and audit trails from day one

Financial institutions ask two questions before features: 'is our data isolated?' and 'is every action logged?'. Retrofitting either into 179k lines is a rewrite. Both were architectural decisions made before the first feature.

6. Token costs are an engineering problem

At enterprise document volume, careless prompting is a five-figure monthly bill. Caching, chunk deduplication, and routing cheap tasks to cheap models cut costs dramatically without touching quality where it matters.

7. One strong generalist can own a serious slice

Orchestration, rendering, exports, and verification — one person, 528 commits. Enterprise AI doesn't always need a big team; it needs people who can hold the whole pipeline in their head.

If you're building an AI product that has to survive compliance review — in fintech, legal, healthcare, or anywhere accuracy is contractual — that's the work I do. Get in touch.