GeniusCFO
An AI accounting workspace for Indian businesses — connect a bank account, ask questions in plain English, and it keeps the books underneath. Schema-first multi-agent system, built from scratch as founding developer and tech lead.
Role
Founding Developer & Tech Lead — Backend & AI
Services
Multi-Agent Systems, LangGraph, Document AI, Hybrid RAG, LLMOps
Year
Ongoing
Overview
GeniusCFO gives Indian businesses a CFO they could not otherwise afford. You connect a bank account and ask questions in plain English — “How much profit did I make last month?”, “When will I run out of cash?” — and the system does the accounting underneath: journals, trial balance, bank reconciliation, balance sheet, notes to account, and a GST workspace with one-click GSTR-3B filing. The whole product runs on a phone. I have been on it since the start as founding developer, and lead the backend and AI architecture.
Indian accounting is not a domain where a confident wrong answer is merely embarrassing. An ITC mismatch that slips through becomes a misfiled GSTR-3B, then a notice, then a penalty three months later. Answers have to be grounded in the business's real filings, traceable to a source, and correct under GST rules — and they have to reach owners who do not know accounting and are reading them on a phone.
Architecture
Approach
Built the supervisor orchestrator: it reads intent, plans a workflow across 9 specialised agents — Extraction, Processing, Retrieval, Reconciliation, Compliance, Reporting, Advisory and more — and dispatches them over a typed agent-to-agent protocol.
Ran long workflows on LangGraph state machines with Redis-backed checkpointing, so a multi-step job survives a restart instead of starting over.
Owned document ingestion end to end for 15+ document types: a four-engine OCR chain (PaddleOCR, Surya, EasyOCR, with AWS Textract as cloud fallback), open-set classification, then per-type schema extraction through forced tool use.
Layered an accuracy harness between the model and the ledger. Deterministic checks — GST arithmetic identities, GSTIN check digits, line-item sums, identifiers grounded against the OCR text — trigger up to two targeted repair rounds, then demote confidence and route to human review rather than post a wrong entry.
Built the conversational layer on hybrid retrieval, dense embeddings fused with BM25 over Qdrant, behind a mandatory per-tenant filter that refuses to run unfiltered.
Gave the assistant long-term memory: idle conversations distil into typed facts, decisions, and preferences that are injected into later chats, scoped per individual so one team member's context never surfaces in another's.
Held the data boundary — for tenants without raw-document consent, PII is tokenised before any LLM call and restored inside the perimeter.
Instrumented the platform with Prometheus, Grafana, Sentry, and LangSmith tracing: metrics at four pipeline seams, queue-depth gauges, and alert rules for worker health and booking-failure spikes.
Built a nightly self-improvement loop — grade, distil, gate, promote — that turns resolved production corrections into versioned, reversible prompt rules, promoted only when a re-run over the golden set does not regress.
Outcome
91.4% field extraction accuracy against the golden document pack, measured on the live path rather than in a notebook.
85% of bank reconciliation lines auto-matched, 10% routed to review, 5% flagged as new entries.
5000+ documents processed a month with no manual step in the loop, and manual audit effort down 90%.
6 microservices in production behind a 7-job CI gate on a protected main branch — no direct pushes, no red merges.
In batched rollout to businesses and partner CA firms, with 150+ businesses in the current cohort.
An agent architecture the team extends without redesigning the core.