Parag Bhagwat
Parag Bhagwat · Dallas · still shipping

I make AI work inside large organisations. And I still build it myself.

  • OrganisationAs Group CDO, put GenAI into an airline, retail and consumer businesses, and migrated a live airline's core system.
  • PlatformAt CGI now: Claude skills, agents and workflows, and an internal marketplace so whole delivery teams can reuse them.
  • Hands-onAt night: a cited RAG over 6,400 pages of ERP docs, a multi-agent comic studio, and the AI you're talking to.
What I've built ↓How this chatbot works ↓The longer story ↓
Parag Bhagwat
Director, Consulting Expert · CGI27 years · 3 startups, 2 acquired · former Group CDO · founder, aicomics.life
Parag AI versionmy digital clone · knows only what I've written
Comic panel: an engineer alone in a dark office at night, April 2026
Panel 1 of Tell Me What Broke. Written by me, drawn by the AI studio I built. ✓✓
Hi, I'm Parag, or the version of me that has read everything I wrote down. Ask about the systems, the airline, the startups, or how I'd deploy AI in your company. Every answer shows its sources, and if I haven't written about it, I'll say so. ✓✓
It only says what I've written down, and code fact-checks it first. Tap "how I answered" under any reply to see the sources. How it works · runs on Claude. Questions are saved so the real me can read them; no names or IP addresses.
What I've built

Systems, with an honest status on each

Problem, what I built, what broke, result. Client names are left out where they aren't mine to share.

Working system · built by me · 2026

A documentation brain for an enterprise ERP

Planentities, typosclarify cards Retrievegraph · lookupsvector + keyword Composebudgeted contextLLM writes, cites Verifyevery ID + pagein evidence?
Problem
41 guides and 6,400 pages, where the answer to one question lives on five pages in three PDFs.
Built
One-time parsing; vision-model descriptions of 1,911 figures; a census of 4,812 tables collapsed into about ten families; a knowledge graph of about 1,000 entities and 970 relationships; hybrid search over 14,000 chunks; four role lenses on one knowledge base; a deterministic verifier.
Broke
Definitions crowded out of context, typos, acronyms that embed poorly, codes only defined in prose. Each became a mechanism plus a test question.
Result
Cited answers in seconds, or an honest "not found", never a guess.
Live · my own product · 2026

A multi-agent AI comic studio

Ashfall, Book One cover, made with the studio Writers Director Art + QC Lettering Motion Release gate 1gate 2gate 3 one "brain" seam: paid API on Telegram · Claude in-session at the desk ($0)
Problem
Turn one person's stories into consistent, multilingual, daily comics without a team.
Built
Eight departments with swappable engines, an episode record as the only contract, shot specs that the director writes, the renderer follows and QC checks, and language as data, so translation edits finished panels instead of redrawing them.
Broke
Fail-open QC, a non-faithful editor, reference order, runaway review spend. What broke →
Result
Two graphic novels in three languages, a daily comic network, 128 commits, about 98 golden tests.
Led as Group CDO · 2025–26 · live and designed

GenAI in an airline, retail and consumer group

Problem
Put AI into the daily operations of people with real jobs to do, across very different businesses.
Built
SGG One AI (a governed prompt gallery, chat over documents and images, prompts as APIs); AI Vigil vision compliance checks across 130–200 stores; a Claude-based in-house dev team (custom skills for BRDs, Claude Code for SAP B1); Copilot for 500+ staff with hands-on training; DusMinute, WhatsApp ordering with AI recommendations for shops inside residential communities (I designed it and led the vendor build; live, 1,000–1,500 orders a month); the SGG One App no-code process engine (designed).
Hardest
Migrating a live airline (Star Air) to the Hitit passenger service system in four months, as acting CTO.
Result
60%+ less manual inspection time · 70% less vendor dependency · 3x faster delivery · up to 40% efficiency gains in finance (internal survey). Figures are from SGG's internal reporting.
Status
Live: SGG One AI (built by an implementation partner to my spec; I owned it), AI Vigil, the Claude dev team, Copilot, DusMinute, the airline migration. Designed, not built: SGG One App.
Client work · 2024–26 · builds, POCs and designs

Agents, knowledge graphs and fine-tuning for clients

Manufacturing
A machine-knowledge system over OEM manuals: a 697-node knowledge graph, hybrid BM25 + vector search, and an error-code chatbot for the shop floor.
Safety
Multi-agent EHS systems: specialist agents over A2A, an MCP tool server, a human-review console; incident capture on Google ADK and Gemini.
AR/XR
When prompts couldn't map intent to engine functions, a fine-tuned model generated executable code straight into the IDE.
Insurance
For an insurtech (Bimakavach): QCR, automated quote comparison reports across insurers written with GenAI, plus plain-English conversation with their database.
QA
Test-case generation with prompt chaining and RAG for a European client: 50%+ faster QA cycles.
What I'm building

On the workbench, at night

Some of these will work. Some won't. All of them are real.

Day job · CGI · now

Claude skills, agents and a plugin marketplace

At CGI I build with Claude: plugins made of skills, agents and workflows, and an internal marketplace so delivery teams can find, reuse and combine them to decompose and automate enterprise delivery work. Not one AI tool, but a way for a whole delivery organisation to use AI safely and repeatably.

That's the public level of detail. Internal names, numbers and clients stay inside CGI.

Live

aicomics.life

Long-form graphic novels built for the phone, plus a daily silent comic network. Behind it is a multi-agent production studio run from Telegram: writers' room, director, art, lettering, motion, with three human gates and about 98 golden tests.

Read Ashfall →
Live · in progress

GreyGenius

AI for India's 45+ professionals. A three-tap AI readiness snapshot, role videos, and a first course where every learner gets their own spend-capped API key.

greygenius.ai →
Launching

NudgeNod

"Did Mom take her meds?" It asks by text, up to three times a day. No app for anyone, and one email each evening. Built, tested (130 tests), in carrier registration.

nudgenod.com →
Being printed

No Passing

A 32-card deck for tech people over 45: "Afraid of AI, or stuck on step one?" Fear, Worth, Gap, Move. You don't get to pass on a hard question.

Proof of concept

Gyan Sessions

15-minute advice calls between young people and elders over 60, with an AI agent on every call listening for scam signals.

Prototype

Sherpa AI

Looks at your screen and tells you the next click in SAP or Zoho, out loud, in Hindi or English. For teams with no training budget.

Prototype

Staked

References the candidate owns: former managers vouch up front, and recruiters open one link.

For the technically curious: case studies, architecture, what broke, and this chatbot's test results.

How this chatbot works

A small, honest RAG, with its test score on the page

The same design as the ERP documentation brain, scaled down to one person. AI only where it's needed; everything else is code you can read.

Planentities · aliasesinjection/privacy guards RetrieveBM25 + vectorsrank fusion Gradeno evidence →"I haven't written that" Composethe only LLM callcites every fact Verifycitations · numbersfail → withheld
Knowledge
About 50 first-person fact cards plus an entity registry with aliases ("Loop" → NudgeNod, "Kaka" → Meter Kaka). If it isn't in a card, the bot doesn't know it.
Pipeline
Plan (intent, entity resolution, typo tolerance, guards for injection, privacy and off-topic) → hybrid retrieval (SQLite FTS5 BM25 + bge-small vectors, fused with reciprocal rank fusion) → grade (no evidence means an honest "I haven't written about that") → Claude writes a short answer that cites its sources → a deterministic verifier checks every citation was retrieved, every number is in the evidence, and no blocked terms appear. Open "Show the trace" under any answer to see every step.
Tests
A golden set written before the code: facts, multi-hop, beliefs, false-premise traps ("Is he still at SGG?"), prompt injection, privacy and off-topic. It runs before every deploy.
—loading the latest run from the live service
Things I've broken

Real failures, and what each one changed

Most portfolios only show what worked. These are real failures from systems I run, and each one became a permanent mechanism and a test.

Broke

The answer that was retrieved, then cut

In the ERP documentation brain, "What is an AD chain?" came back "not found", even though the definition had been retrieved. One role's evidence ordering put text last, and the context limit cut it off.

Changed

Budgets, not orderings

Context now reserves a minimum budget per evidence type, so a definition can't be crowded out. The question became a permanent eval case, and every role is re-tested after every change.

Broke

The QC that approved its own bad panels

When the reviewer model returned malformed JSON, my verifier quietly accepted the panel "unchecked". It failed open, and I found out from production logs.

Changed

Fail closed, and prove it

Strict parsing, a QC that blocks rather than passes when it's unsure, and tests that deliberately break the checker to prove it can fail.

Broke

An "edit" that redrew everything

Asked to swap a few words in a panel, one image model regenerated the whole thing. I measured 26–42% of pixels changing on what should have been a text edit.

Changed

Measure before you trust

Text swaps go only to an editor that was verified pixel-faithful side by side, and every batch gets eyeballed before spend.

Broke

Continuity, silently

A location reference was attached after the previous panel, so the model read the location as "the previous panel". Characters drifted and nothing errored.

Changed

Order is a contract

Reference order is now specified and tested: cast sheets first, location next, previous panel always last.

Broke

The $100 that disappeared

Paid review calls quietly ate a big share of a credit top-up before anything looked wrong.

Changed

Pay only for pixels

One "brain" layer serves both the paid API and Claude in-session for $0. Same prompts, same checks. Only image generation costs money.

Broke

Great POCs, no customers

At RazorSwift we built resume search over 100,000+ resumes and agents on top. The pilots worked. We didn't convert them.

Changed

Start with "for whom"

I now scope with the people who'll use it before writing code, and I treat a POC as a sales instrument, not a finish line.

Things I believe

About AI, and about building

Build where people already are.

In India that usually means WhatsApp, and I've been building AI there since 2020: YAYSHOP let 15,000+ sellers run their shops in chat (acquired) · RazorSwift built profiles from WhatsApp conversations (still live) · DusMinute takes 1,000–1,500 orders a month by WhatsApp with AI recommendations (live) · NudgeNod does daily check-ins by text (launching). No app to install, if a message will do.

Efficient, observable, connected. Then autonomous.

The mandate I hear most is "eliminate X with AI." You can't automate a process nobody can see. Make it efficient, then observable, then connected. Autonomy on top is the easy part.

Writing code with AI isn't the problem.

Writing code without structure is. Every question goes straight into a model now, and nobody asks the user first. Those code bases will be very hard to maintain in three years.

Know why, and for whom.

The lesson from 2003 still holds. If you can't say who it's for, the model can't save you. RazorSwift taught me this the expensive way.

AI only where it's needed.

Deterministic code for what must be exact; the model for language. Every answer should be checkable, and AI failures should degrade to a human, not to a wrong answer.

Most companies get stuck after the demo.

Making AI real is a different skill from showing it. My own projects climb the same ladder: a one-off, then repeatable, then a system.

Experience isn't obsolete.

It's what you need when things break. AI can build the app. It won't think the way you would.

The longer story

The person behind the systems

Eight short chapters, in my words: ships, Y2K, leaving at 45, a pivot that was really survival, the startup that taught me most, an airline, and comics at night.

011996–2003

From ship couplings to code

I'm a mechanical engineer. My first job was fitting couplings and pumps, and I used to go out on ships to do it. I come from a middle-class family where everyone had a steady job and nobody had built a business. My parents gave us everything we needed, and I'm proud of where I come from.

Then Y2K happened and everyone was going to America through software. So I taught myself Java. Nobody would hire a mechanical engineer with no experience. Then a friend starting his own company gave me a chance, and I learned PowerBuilder on the job. After that came three years as a contractor to American Management Systems, until AMS hired me in 2003 and CGI bought AMS.

022003–2019

Sixteen years, one lesson

At CGI I became Chief Technical Architect on an ERP platform that US state governments run on: $10M+ modernizations for Texas agencies, under strict compliance, where clever code gets punished and boring code survives. In 2017 I moved to Bangalore and set up CGI's Emerging Technologies Group: AI, chatbots, cloud.

“CGI taught me that not everything is just code. You have to know why it's being written, and for whom.”

It also taught me that process matters. As an engineer I hated process. Then my team grew. And you can stay at one company as long as you want if you have good friends and mentors. I had both.

032019

Leaving at 45

I was senior and comfortable. But I'd always wanted to build something of my own, and nobody in my family ever had. I tried a couple of interviews and couldn't land anything: they wanted B2C, and my experience was B2B. One interviewer told me my resume was good, but he didn't think I could build something that scales.

So I went and built it. Almost everybody thought I was crazy. Friends asked if I really had to leave. My wife backed me and went back to work so we could do it.

042019–2022

They could wait. We couldn't.

TeamKaro was group buying: prices drop as more people join. We launched in March 2020 with a plan for pop-up shops in office complexes and stadiums, and we were doing 500+ orders a month. Then COVID shut everything.

“Investors said they'd wait till COVID blew over. They could wait. We couldn't.”

My co-founder and I had given ourselves two years and had limited money left. Two founders, one developer. We asked what problem was real right now, and rebuilt what we had to run on WhatsApp. That became YAYSHOP: 15,000+ small sellers running their catalogue, orders and payments through chat. The pivot wasn't strategy. It was survival.

We went 22 months without a salary. When Infra.Market acquired both companies in 2021, it wasn't a victory lap. It was relief. Stay alive, so you can go to battle once more. I stayed on as Director of Engineering, leading 50+ engineers through the integration.

052023–2025

The one that taught me most

RazorSwift was going to bring hiring online at scale. We built our own models before GenAI, rebuilt on OpenAI's API when it arrived, and watched resume parsing become trivial overnight. So we went after a bigger problem: recruiters sitting on hundreds of thousands of resumes and still paying to find candidates they already had. We built a natural-language search over 100,000+ resumes, plus agents that chase missing details.

“We could build anything. We couldn't sell it. POCs aren't customers.”

That's the honest version. The pilots went well, but we didn't convert them. I stepped back in early 2025 for family. Alongside it I'd been advising startups and enterprises in India, Europe and Canada on getting AI from POC to production: fine-tuning, RAG, agents. The pattern was always the same. Everyone knew they needed AI. Nobody knew how to make it real.

062025–2026

A founder took a Star Air flight

I got my next job because another founder at my co-working space flew Star Air, met the managing director, and later saw that the group needed a tech leader. He connected us. That's how I became Group Chief Digital Officer (CDO) of the Sanjay Ghodawat Group: an airline, retail, consumer goods, education, all at once. I led three business CTOs.

The hardest thing I did there was also what I'm proudest of: migrating a live airline's reservation system (Star Air, to the Hitit passenger service system) in four months, while I was acting CTO. Planes don't stop flying while you migrate.

The rest was putting AI into the daily work of people with real jobs to do: store staff, airline crews, finance teams. We launched SGG One AI, a governed prompt gallery where a good prompt stops living in one person's chat history and becomes something the whole group can use. We used vision AI to check stores and aircraft against a standard. I built an in-house development team on Claude, and I put Copilot in front of hundreds of people, with training rather than just licences. The numbers are in the evidence.

I'm grateful to SGG. They gave me an airline to migrate.

072026

Nights

I grew up on Indrajal Comics: The Phantom, Mandrake, Bahadur. People have always told me I think differently. Now there's a machine that can draw what I think.

Ashfall came from a question I'd carried for years. Everyone tells the story of the Mahabharata war. What about the people 500 miles away from it? Tell Me What Broke is closer to home. In Indian tech there's an unwritten rule that you're done at 50. It's about an engineer let go in the AI layoffs, then called back because he's the only one who understands why the system was built the way it was.

To make them I built a studio that runs from Telegram. Claude writes, directs and checks every panel; image models draw; I approve at three gates. Now aicomics.life publishes long-form comics built for the phone, plus a daily silent comic starring Meter Kaka, an honest Mumbai auto driver.

Read Tell Me What Broke →   Read Ashfall →

08Now →

Back where I started

In 2026 I moved to Dallas for my family, and went back to CGI as Director, Consulting Expert, bringing AI into the same kind of enterprise systems I used to architect. It closes a loop.

A normal Tuesday: days for CGI, nights for calls with India, and everything in between I build. My family is in India. We talk every day.

“At 70 I want to hand my daughters an empire: hospitals, a school, with my forefathers' names on them.”

I've got at least twenty more years of building in me.

Timeline
  • 1996BE Mechanical, University of PuneThen couplings, ships, and Java at night
  • 1998–2003Vikalp Systems → Prescient (contractor to AMS)PowerBuilder, then enterprise consulting
  • 2003–2019CGI, USA and IndiaChief Technical Architect, then Head of Technology; Emerging Technologies Group
  • 2019–2021Co-founder & CTO, TeamKaro → YAYSHOPAcquired by Infra.Market, 2021
  • 2021–2022Director of Engineering, Infra.Market50+ engineers, OMS, payments, delivery
  • 2023–2025Co-founder & CTO, RazorSwiftPlus independent GenAI architect for clients in India, Europe and Canada
  • 2025–2026Group CDO, Sanjay Ghodawat GroupActing CTO, Star Air
  • 2026 →Director, Consulting Expert, CGI · DallasFounder, aicomics.life
What colleagues said
“One of the most productive people I have ever encountered… He quickly earned our customer's trust.”Don Cartmell, who managed me on a project (LinkedIn)
“Always calm and controlled and capable of finding a solution whenever one is required no matter how hard the problem is.”Sami Boudriga (LinkedIn)

Download resume (PDF)

Talk to me

No pitch. No invoice.

Free conversations about founding in the messy middle, AI that won't leave the demo stage, or the second half of a tech career. I don't take paid consulting while I'm at CGI.

Email me LinkedIn WhatsApp
Yes or noNo ghosting
< 24 hrsOn weekdays
HonestEven if it's "no"
Message me