Nilay Chindaliya, Mumbai · Founder's office, IndiaFilings · AI & DS '26

AI Engineering
×Financial Markets
×Compliance Infra

I build operational intelligence, software that doesn't just inform an operator, it changes what the operator does next. Right now that means AI agents talking to real customers of India's largest compliance platform, and analyst terminals built from real filings with every figure cross-checked by arithmetic.

4 AI agents live 500+ real customer conversations 8 live systems one click away 5 transcription errors caught by math alone --:--:-- IST · at the desk, Mumbai

How I got here

Every builder has a first system. Mine was a shop counter.

Chapter 01The counter

I grew up inside a family business, not next to one.

Retail, wholesale, manufacturing, an NGO, a law firm, a school, all around one dinner table. From the counter I learned how cash actually moves through a small business, and how trust is the real currency of Indian commerce. During lockdown I scaled the family's Amazon FBA business roughly 10x on pricing, demand and inventory cycles. Before any degree, I'd learned the lesson I now see everywhere: things that exist are not systems that connect.

Chapter 02The banking education

A senior ICICI banker gave me a curriculum instead of an internship.

Two months, three hours a day: how banks earn, ratio analysis on real annual reports, bank audit, trade finance, treasury. The sentence that reorganised my thinking, compliance is risk, is banking. A missed GST filing isn't a clerical event; it's a credit-risk signal. He then steered me away from prestige finance and toward fintech, where AI skills and business instinct compound instead of compete.

Chapter 03The operating floor

At IndiaFilings I audited where the service layer actually breaks.

Working from the founder's office, six days a week: a churn audit across 50+ on-hold GST customers, a 415-lead pipeline audit with per-advisor CRM discipline scoring, a revenue-integrity investigation for the founder, and a 200-account reassignment audit on the international desk. The finding underneath all of it: humans were drowning in exactly the work agents are built for, discovery, classification, documents, follow-up.

Chapter 04The build

So I built the systems. Live. On real customers.

Four agents live at IndiaFilings, one iterated v1 to v12 against real conversations, with a 50-test eval suite and an LLM-as-judge pipeline scoring every live conversation. A financial command centre verified against a live business carrying 5,163 invoices. And the terminals below, built from real filings. Every one exists because an audit I personally ran demanded it.

The systems

10 BUILDS · 4 CATEGORIES

Everything here runs for real users or was built from real filings and live data. Eight are one click away. The rest are worth the call.

01 · Investment Analytics & Terminalsdocuments → instruments
02 · Live Operationslive data · real stakes
03 · Compliance & Accountinglaw, encoded
04 · Personal Systemsmy own problems first

    

The method

"Every financial document is a database wearing a PDF costume."

01

Extract & structure

Prospectuses, filings, invoices, CRMs, structured data flattened into dead pages. I pull it back out into typed, connected form.

02

Verify by arithmetic

Balance-sheet identities, CAGRs, dilution math, penalty accruals, cross-checked, not vibes. AI classifies; it never invents a number.

03

Ship to real users

One feature end-to-end, validated on live data, then the next. "Shippable" means one real business depending on it, live.

The through-line: India runs on 60M+ MSMEs whose financial reality is trapped in ledgers and filings that don't talk to each other. I'm building the layer that lets them see themselves.

Scar tissue, formalized

What the builds taught me

Every system on this page left a scar or a principle. These are the principles.

From AI × financial systems

  1. LLMs classify; arithmetic verifies. A model never gets to invent a number, the Jio errors were caught by balance-sheet identities, not by the AI.
  2. Eval suites beat clever prompts. July moved v1→v12 on a 50-test suite and an LLM judge, not on prompt tweaking.
  3. Citation discipline must be structural, not instructional. Telling an agent "don't hallucinate" does nothing; restricting it to cite only what retrieval returned does everything.
  4. Production APIs are what they are, not what the docs say: typo'd fields, inconsistent pagination, dual-call auth. Live means handling the mess.
  5. 500 real conversations teach what no dataset can: customers don't ask questions in your taxonomy's order.
  6. Compliance is credit risk in disguise, a missed GST filing is a risk signal, not paperwork.
  7. An agent that says "I don't know, escalating you" keeps the customer; one that guesses loses them.

From the operating floor

  1. Dissatisfaction is detected too late by design: the refund request is the last signal; repeat calls, silence, and blown timelines were already sitting in the CRM, unread.
  2. Most service holds are communication failures, not technical ones.
  3. A ticket without one named owner is a ticket that ages.
  4. Escalation needs a clock, not a queue: day-count ladders, each rung a different desk.
  5. CRMs record activity, not health; scoring per-advisor discipline exposed the gap between the two.
  6. Humans were drowning in exactly what agents are built for: discovery, classification, documents, follow-up.
  7. An audit earns the right to build. Every system on this page exists because an audit I personally ran demanded it.

The notes from each build get formalized into reference volumes I keep, the Agents Master Book, the Banking Codex, the Field Guide. Writing forces the discipline reading skips.

The operator layer

The commercial instinct

Before I wrote code, I ran a counter. Retail, wholesale, manufacturing around one dinner table: acquisition, pricing and retention weren't subjects, they were dinner conversation.

  1. Acquisition starts with trust, not funnels: at a counter, the repeat customer is the business model.
  2. Pricing is a live experiment. Scaling the family's Amazon FBA ~10x during lockdown was daily repricing against demand and inventory cycles, not a strategy deck.
  3. Inventory is cash wearing a costume: working-capital instinct before I knew the term.
  4. Word of mouth is the only channel that compounds; every other channel you rent.
  5. Innovation is removing one step of friction for the customer, not adding a feature.
  6. You find demand by listening for what people ask for and can't get: the same instinct behind every audit I've run since.

The work I keep choosing

Roles where AI engineering meets financial operations meets ownership, embedded in a real business, building against live data. The forward-deployed pattern.

  • Applied AI / GenAI engineering, in banks and financial institutions
  • Forward-deployed / AI solutions engineering, embedded with the operation
  • Fintech founding-team & founder's-office roles, own the loop end-to-end
  • Financial data & analytics engineering, data exhaust into instruments

I optimize for live business data over sandboxes, ownership over ticket queues, and proximity to the P&L. The systems above are the interview.

Get in touch

Building financial infrastructure? Let's talk.

I optimize for live business data over sandboxes, ownership over ticket queues, and proximity to the P&L. If your team works that way, we should talk:

© 2026 Nilay Chindaliya · Mumbai /jio · /zepto · /tcs · /oi · /lexos · /hni · /hni/tool · /agents · /creditos · built as a system, like everything else here