From business problem to AI system in production.

I'm Anubhav. I sit with a business, find the problem worth solving, and ship the AI system that fixes it, tested end to end.

CEO & Lead Engineer, Toorunt AI · Co-founder, CEO & Lead Engineer, Neural Weaver LLP · Mumbai

I sit with the client,find what's broken,and ship the fix.

(01) About

I bring the finance depth to understand the problem and the engineering depth to ship the fix, and I have led each of my startups as both CEO and lead engineer.

Now

  • Toorunt AICEO & Lead Engineer
  • Neural Weaver LLPCo-founder, CEO & Lead Engineer

Before

  • Axis BankSenior Manager, Business Analytics
  • IIFL CapitalAssistant Vice President, Research Analyst
  • HeltarLLM sales agent for Spinny

Education

  • IIT KharagpurB.Tech Biotechnology + M.Tech Financial Engineering
  • Codeforces ExpertRated 1600+

(02) Work

Thirteen projects, six told in full

Pick a kind of work to narrow the list. Six rows open a full case study below, and three have a film you can watch on this page.

13 shown
  1. 01 Toorunt AIAn AI engineering team that ships a company's Jira backlog inside its own Jira and GitHub, with people signing off at three points. Own startupCase studyFilm 4:0711,153 tests · 0 failed
  2. 02 NaviGoInA full ride-hailing platform for a Mumbai start-up: rider app, driver app, fleet console and backend. Client deliveryCase study574 Go + 249 browser tests
  3. 03 MSETCL Innovation PortalA secure, bilingual innovation portal for Maharashtra's state electricity transmission company. Client deliveryCase study347 / 347 test cases
  4. 04 CA AutomationSoftware prepares the books and the compliance work for CA firms. The chartered accountant approves. Own productCase studyFilm 4:4617 banks · 1,100+ tests
  5. 05 Pump LedgerWhatsApp meter photos replace manual shift-end reconciliation at a working IndianOil fuel station. Client deliveryCase study43 / 53 auto · 0 wrong
  6. 06 OutboundAn AI sales agent where models draft and judge, a person approves, and policy decides what may be sent. Internal toolCase study1,600+ tests
  7. 08 The Market SageEquity research over Indian company filings. Its research reviewer reads broker reports, stages cause-and-effect links, checks each with a blinded AI reviewer, and merges only what a person approves. Own platform470,683 NSE filings · 499 companies
  8. 11 Pharma distribution ERP rebuildReverse-engineered a Windows ERP that stored its data in shared DBF files, then rebuilt its daily core on a transactional server with the original screens and hot keys. Legacy modernisation302 DBF tables mapped · 48 reports
  9. 12 Delivery-signal studySix attention signals from NSE delivery data on large caps, tested walk-forward over 2016–2026. The extra filters did not beat the baseline out of sample, and the report says so. Independent research2022+ out of sample: 20.3% a year · Sharpe 0.82
  10. 13 Film production toolkitTurns a brief into a finished product film: script, narration in English, Hindi or Marathi, motion graphics timed to the voice, an original score, captions, and loudness and accessibility checks. ToolingMade five of the films on this page

(03) Case studies

Six systems, told in full

Each one: the problem, what I built and the result, with the proof underneath.

Case 01 / 06Own startup · CEO & Lead Engineer · 2026

Toorunt AI

An AI engineering team that ships a company's Jira backlog inside its own Jira and GitHub, with people signing off at three points.

Problem
Most teams have more tickets than engineers. AI coding tools can write the code, but nobody can say who approved a change or prove it was tested before it shipped.
What I built
A platform that takes each ticket from plan to release. People sign off only on the plan, the pull request and the merge. Every change clears 14 deterministic gates, each bot has its own least-privilege Jira and GitHub identity, review is bot-to-bot and adversarial, and every decision lands in a hash-chained audit log.
A contract compiler turns a signed PRD into acceptance tests, and each test must fail before any code is written. Builds run in distributed work packets with Redis leases and fencing tokens. Code is built in one place and judged in another: CodeBuild makes the image, and credential-less Fargate sandboxes verify it.
Result
A ticket on a demo codebase went from intake to deployed release in 2 h 51 min. Release CI passed 11,153 tests with 0 failures. A staging run of 4 work packets on 3 workers reached 100% first-pass success.

Follow ticket KIRA-24

One ticket, intake to release

BotworksGatechecksPerson signsdecides
  1. 01Bot
    Intake

    A ticket lands in the customer's own Jira. A bot with its own least-privilege identity picks it up.

  2. 02Person signs
    Plan

    A planning bot writes the plan. A person signs it off. Sign-off 1 of 3.

    KIRA-24: intake to plan in 21 min 41 s

  3. 03Gate
    Contract

    The signed PRD compiles into acceptance tests. Every test must fail before any code is written, and passes only on the exact candidate revision.

  4. 04Bot
    Build

    Work splits into packets that run on distributed workers, with Redis leases, fencing tokens and work stealing. CodeBuild makes the image.

    KIRA-24: plan approval to pull request in 1 h 31 min

  5. 05Gate
    Verify

    Credential-less Fargate sandboxes judge what was built, separately from where it was built. The change has to clear 14 deterministic gates.

  6. 06Person signs
    Review

    A separate bot reviews the pull request adversarially. Then a person approves it. Sign-off 2 of 3.

    KIRA-24: pull request to peer approval in 9 min 27 s

  7. 07Person signs
    Merge

    A person approves the merge. Sign-off 3 of 3. Organisation policy can pin these approvals so a team can only tighten them.

  8. 08Bot
    Release and watch

    The release ships and is watched after merge. Every decision on the way is written to a hash-chained audit log.

  9. Released

    2 h 51 min

    From intake to deployed release for KIRA-24, on a demo codebase.

    Release CI: 11,153 passed · 0 failed

Case 03 / 06Client delivery · via Creant Technology · Oct 2026Delivered for client review

MSETCL Innovation Portal

A secure, bilingual innovation portal for Maharashtra's state electricity transmission company.

Problem
MSETCL wanted employees to submit ideas and committees to evaluate them fairly, in Marathi and English, on Liferay DXP and to a written SRS with government-grade security.
What I built
Three custom Liferay modules: two-factor sign-in, idea submission, two-stage blind committee scoring, MIS reports and a hash-chained audit trail. Every upload is scanned by ClamAV and re-rendered through Gotenberg before anyone opens it.
The build targets OWASP ASVS Level 2 and India's DPDP Act 2023.
Result
Delivered with 347 of 347 test cases passing, 136 of them security attacks. Load test: 94 requests a second with 0 errors and a 0.10 s dashboard p95. A restore drill rebuilt 264 documents. The handover included a 45-page delivery report.

347 test cases, all passing

Security suitesFunctional suites

Test cases by suite, from the delivery test workbook.

How an idea moves

  1. 01
    Sign in

    Employees sign in with a second factor and use the portal in Marathi or English.

    Employee
  2. 02
    Submit the idea

    The idea and its attachments are submitted through the portal.

    Employee
  3. 03
    Scan every upload

    ClamAV scans every uploaded file.

    Check
  4. 04
    Re-render documents

    Gotenberg re-renders each document, so reviewers open a clean copy instead of the original upload.

    Check
  5. 05
    Blind scoring

    Committee members score each idea without seeing who submitted it.

    Committee
  6. 06
    Second stage

    Scoring runs in two stages before an idea is ranked.

    Committee
  7. 07
    Ranking and MIS

    Rankings and MIS reports come from the recorded scores.

    System
  8. 08
    Audit trail

    Every action is written to a hash-chained audit trail that can be verified end to end.

    System
Case 04 / 06Own product · Neural Weaver LLP · Oct 2026 – nowIn testing at a practising CA firm

CA Automation

Software prepares the books and the compliance work. The chartered accountant approves.

Problem
At month-end, CA firm staff re-key bank statements and bills into Tally by hand, then reconcile GST and TDS line by line. It is slow, and a missed item becomes a penalty.
What I built
Statements from 17 Indian banks become draft vouchers, including scanned and password-protected PDFs. Client rules decide first and an LLM second, OCR is checked against running balances, and every correction becomes a rule, so a mistake is fixed once. Staff see only the exceptions.
Approved entries go to Tally through an idempotent XML push. The same review-first flow covers GST returns and 2B matching, TDS under the old and 2025 Acts, capital gains, PF/ESI and MSME 45-day dues, behind an append-only audit trail and period locks.
Result
Scoped from recorded Tally walkthroughs and a competitor teardown, now in testing at a practising CA firm, with 1,100+ automated tests.
A scanned statement becomes ready entries, each matched to a party and the bill it settles. 10 s from the film, muted.
CA Automation review screen: a bank payment with a suggested ledger, the reason for the suggestion, and the source statement with balances tying out.
The review queue shows only the lines that need a person. Each suggestion says why it was made, and balances tie out on every row of the source statement. Demo company.
CA Automation GST screen: supplier invoices matched against GSTR-2B and IMS, with counts for matched, value differs, on portal but not in books, and in books but not on portal, and the credit going into GSTR-3B.
Supplier invoices matched against GSTR-2B and IMS. Each mismatch says what to do next, and the credit going into GSTR-3B is worked out. Demo company.
Case 05 / 06Client delivery · Heera Associates, IndianOil dealer, DombivliLive since 2 Oct 2026

Pump Ledger

WhatsApp meter photos replace manual shift-end reconciliation at a working fuel station.

Problem
Pump staff already sent meter photos on WhatsApp at every shift change. The owner still matched fuel sold against cash, UPI and credit by hand, and shortages surfaced days late.
What I built
A custom offline reader for the dot-matrix displays on fuel dispensers. Each photo is read twice by independent methods, and a reading is accepted automatically only when both agree. Everything else goes to a review queue.
A Hindi/English staff app, a daily owner digest that flags cash shortages and stock variances, and TallyPrime export. Nothing touches the dispenser hardware, in line with oil-company rules, and data handling follows the DPDP Act.
Result
On a 53-reading benchmark, 81% were accepted automatically with zero wrong. The owner has run it on real data since 2 October 2026.

Two reads must agree

Read A1234.56
Read B1234.56
Both reads agree. Posted to the ledger automatically.

Illustration with made-up values. The rule is the real one: no agreement, no automatic posting.

Benchmark: 53 readings, one square each

43accepted automatically
10sent to a person
0accepted wrongly

13 real photos and 40 synthetic readings.

Case 06 / 06Internal tool · Neural Weaver LLP · Nov 2026Dry run before launch

Outbound

An AI sales agent where models draft and judge, a person approves, and policy decides what may be sent.

Problem
A young firm needs outbound sales at near-zero cost, without breaking anti-spam or privacy law in any country it writes to.
What I built
An agent that finds companies in open map data, reads their websites, scores them and drafts an email. One Claude model drafts and a different model judges each draft, then a person approves it. Every send passes a single policy gate with country rules for the US, Canada, the UK, the EU, Australia and India, a consent ledger and a suppression list.
The AI may stop contact with someone on its own but can never start it. Sending runs through an exactly-once outbox, my own RFC 3464 bounce parser, RFC 8058 one-click unsubscribe and deliverability circuit breakers.
Result
On the first run, open listing data found a contact address for 342 of 515 companies, against 50 of 116 from websites alone. More than 1,600 automated tests cover the system ahead of launch.

Who decides at each step

From a map listing to a sent email

AutomatedAIPersonPolicy
  1. 01Automated
    Find companies

    A weekly finder searches open-licence map data (Overture Maps Places).

    First run: contacts for 342 of 515 companies from listings, 50 of 116 from websites alone

  2. 02Automated
    Read the website

    It reads each company's own site. Contact data taken from LinkedIn is refused.

  3. 03Automated
    Score fit and size

    Each company is scored for fit and size before any email is drafted.

  4. 04AI
    Draft the email

    Claude Opus drafts each email through the Message Batches API.

  5. 05AI
    Judge the draft

    A different model, Claude Sonnet, judges the draft as a QA gate. The model that writes never grades its own work.

  6. 06Person
    Approve or reject

    A person approves the email before anything can be sent.

  7. 07Policy
    Check the law

    One can_send() gate applies country rules for the US, Canada, the UK, EU states, Australia and India (DPDP), a consent ledger with ISO/IEC TS 27560 fields, and keyed-hash suppression.

  8. 08Automated
    Send once, then listen

    An exactly-once outbox sends through Gmail, Graph or SMTP. Bounces go through my own RFC 3464 parser, unsubscribes through RFC 8058 one-click links on a Cloudflare Worker, and circuit breakers stop sending if deliverability drops.

The AI can stop contact on its own. Only a person can start it.

(04) Screening room

Product films

I write and produce short films that explain a product to the people who will buy it: script, voice, motion graphics and real product screens. Pick a film, then jump to any chapter.

Toorunt AI

An autonomous AI engineering team that takes a ticket from plan to production: 14 gates, three decisions kept by people, and on call after launch.

Product walkthrough on real product screens with test-mode data.

Chapters

    (05) Before my startups

    Earlier

    1. May–Jul 2026

      IIFL CapitalAssistant Vice President, Research Analyst

      Institutional equity research.

    2. 2023–2026

      Axis BankSenior Manager, Business Analytics

      Promoted on the fast track. Owned analytics for a Rs 70,000+ Cr portfolio. Built a CIBIL-trigger pipeline that turned live bureau enquiries into frontline leads (Rs 2,000+ Cr in new sanctions) and a straight-through renewal workflow that cut renewal turnaround by 75%.

    3. 2023–2024

      HeltarWhatsApp sales agent for Spinny

      An LLM agent with tool calling and controlled context routing handled 10,000+ leads. Test drives rose from about 5 to 100+ a day, and prompt compaction cut the cost per test drive from Rs 200 to Rs 90.

    4. 2018–2023

      IIT KharagpurB.Tech Biotechnology + M.Tech Financial Engineering

      Dual degree. Codeforces Expert (1600+). First place at the Inter-IIT Social and Cultural Meet 2019 with the IIT Kharagpur contingent.

    (06) Contact

    Have a problem that needs to reach production?

    I'm looking for forward-deployed and applied-AI engineering roles where the job is to understand a customer's business and ship the system that fixes it.