Luca:
ICAI CA-GPT
A ~199,000 LOC TypeScript monorepo built for Chartered Accountants in India — multi-provider LLM orchestration, pgvector-backed RAG, 11 chat modes and 10 statutory-compliance modules over 106 PostgreSQL tables. Part of the FinACEverse platform.
Commercial product — the repository is private.
CORE CAPABILITIES
Multi-Provider LLM
Anthropic Claude, OpenAI, Google Gemini, Azure OpenAI and Perplexity behind one registry with health-scored fallback routing.
Health Monitoring
A health monitor tracks provider availability so the orchestrator can reroute traffic.
pgvector RAG
Document ingestion into a pgvector store, with semantic triage over retrieved context.
Conversation Memory
Persistent memory and a continuous-learning loop across sessions.
Statutory Modules
Ten India compliance modules spanning GST, MCA, Companies Act and Income Tax workflows.
106-Table Schema
A single Drizzle schema of 106 PostgreSQL tables backing the whole platform.
Product Interface
Assistant, compliance modules and analytics in one workspace.
Multi-Mode Chat
MIS Reports
Conversation Analytics
Product Landing
The routing layer
Five providers, one registry
A compliance assistant that stops working because one vendor is rate-limited is not usable during a filing deadline. So no part of the application talks to a model vendor directly. Every call goes through a registry that picks a provider, and a health monitor that decides which providers are currently allowed to be picked.
5 LLM PROVIDERS
claude.provider.ts
Anthropic Claude.
openai.provider.ts
OpenAI.
gemini.provider.ts
Google Gemini.
azureOpenAI.provider.ts
Azure OpenAI — the guaranteed ultimate fallback.
perplexity.provider.ts
Perplexity, for retrieval-shaped questions.
Plus registry.ts and healthMonitor.ts.
Voice runs the same pattern with its own four providers — Azure, Deepgram, ElevenLabs and OpenAI — so speech is swappable on the same terms as text.
HOW FALLBACK IS DECIDED
Every provider carries a health score
healthMonitor.ts holds a 0–100 score per provider. Successes raise it, failures lower it, and the size of the penalty depends on what went wrong — a rate limit costs more than a single transient error, and an auth or quota failure zeroes the score outright.
Cooldowns are typed, not uniform
A rate-limited provider is parked until a cooldown expires rather than retried immediately. Different failure classes get different cooldowns, and a provider recovers score when its window clears.
Below the threshold, traffic reroutes
A provider under the health floor is skipped by the registry entirely, so a degrading vendor stops taking requests before the user notices it.
Azure is the floor
One provider is always reachable. When every preferred route is unhealthy the orchestrator lands on Azure OpenAI, so the failure mode is a slower answer rather than no answer.
RETRIEVAL & MEMORY
pgVectorStore.ts
The embedding store, sitting in the same PostgreSQL instance as the relational data.
documentIngestion.ts
Chunking and embedding of uploaded client documents.
semanticTriage.ts
Decides what retrieved context is worth putting in front of the model.
conversationMemory.ts
Persistent memory across sessions rather than per-thread context only.
continuousLearning.ts
Feeds observed conversation outcomes back into the system.
Embeddings live in pgvector inside the same PostgreSQL database as the 106 relational tables, so a retrieved chunk and the engagement record it belongs to are one query apart rather than two systems apart.
Clarify first
The assistant is allowed to not know yet
Accounting answers are conditional. The treatment of a transaction turns on facts the question rarely carries — which framework applies, which year, whether the entity is a company, what the counterparty is. A model that guesses those silently produces something that reads correct and is not.
So the assistant writes two blocks into its own answer: an Unconfirmed — answer these to refine list, and an Assumptions block stating what it proceeded on. The orchestrator parses those questions back out of the response and gates the next turn on them, so when the user replies to only some of them the system knows which are still open instead of treating the reply as a fresh question.
The effect is that uncertainty is visible in the output rather than hidden inside it — the user can see exactly which assumptions the answer is standing on.
Spreadsheets as a compiled artefact
The Excel subsystem does not let a model write cell formulas. The model emits an intermediate representation that names inputs, rows and scenarios, and a deterministic compiler turns that into the workbook — picking cell addresses, sheet layout and named ranges itself. Because the IR has no cell-address surface at all, a whole class of cross-sheet reference bugs becomes unrepresentable rather than merely discouraged by a prompt.
11 CHAT MODES
A mode is not a prompt preset. Each one changes the shape of the output as well as the reasoning behind it — Checklist returns a worked checklist, Audit Plan returns a plan, Calculation shows the working, Scenario Simulator varies inputs, Deliverable Composer produces a client-ready document, Roundtable runs several viewpoints against each other, and Spreadsheet emits a workbook. Deep Research and Forensic Intelligence widen the retrieval pass before answering. Picking the mode is how a user tells the system what kind of artefact they actually want back.
10 INDIA COMPLIANCE MODULES
Tally Sync & Finalization
Ledger sync out of Tally into the year-end finalization workflow, so the books a firm already keeps become the starting position rather than something re-entered.
GST Reconciliation
GSTR-1 and GSTR-3B set against the purchase register, with input tax credit mismatches flagged where the return and the books disagree.
Working Paper Library
Reusable audit working papers held centrally, so a firm's house format survives between engagements and between staff.
Fixed Assets Register
Dual compliance on one asset base — Companies Act Schedule II straight-line depreciation alongside the Income Tax block-of-assets written-down-value method, with section 50 capital gains on disposal.
Balance Confirmations
External balance confirmations under SA 505, sent to counterparties as OTP-gated public confirmation pages so a third party can respond without an account.
MCA Filings
AOC-4 preparation for filing with the Ministry of Corporate Affairs.
CMA & Projections
Credit monitoring arrangement data and projections, computing maximum permissible bank finance under Tandon Method II and the Nayak Committee formula.
MIS Reports
Management packs — profit and loss, balance sheet and the key ratios drawn off them.
FEMA & Transfer Pricing
Form 15CA and 15CB workflows for foreign remittance, backed by a remittance register.
Tax Computations
Statutory tax computation workings carried through to the filed position.
Lineage
Three generations
Luca is the third build of the same idea, and the schema records it: 73 tables, then 104, then 106. The middle generation is where most of the domain work happened; the third was mostly a re-homing.
01 — ICAI-CAGPT
Apr–May 2026 · 73 tables
The first build. A chat assistant for Indian Chartered Accountants with its own self-hosted authentication and admin, and the first version of the multi-provider orchestration.
02 — LucaAgent
Jun 2026 · 104 tables
The generation that did the real work. The entire India statutory stack — GST reconciliation, fixed assets, balance confirmations, CMA, MCA filings — was built here, which is where the table count jumps.
03 — Luca
Aug 2026 · 106 tables
Re-homed into the FinACEverse platform shell. The trade was deliberate: the self-hosted auth and admin surfaces were given up in exchange for platform integration, and the domain work carried over intact.
Rewriting twice was not free, but each move dropped something that had stopped paying for itself — first a narrow chat surface, then a bespoke auth and admin stack that the platform already provided. What survived all three generations is the part that was hard to build: the statutory modules and the provider routing under them.
TECHNICAL ARCHITECTURE
frontend
React
Client of the TypeScript monorepo
TypeScript
~199,000 LOC end to end
Multi-Mode UI
11 chat modes in a single surface
backend
Express
API layer of the monorepo
PostgreSQL + Drizzle
106 pgTable definitions
pgvector
Embedding store behind RAG
ai
Provider Registry
Claude, OpenAI, Gemini, Azure OpenAI, Perplexity
AI Orchestrator
Health-scored fallback routing
RAG Pipeline
Document ingestion + semantic triage