Utkarsh Booking July '26
Resume / 01 Long form

Utkarsh Agarwal

Currently AI PM, Cars24
Local BLR /
Status Booking - 1 slot, July 
AIRO ----> 50+ voicebots
CLIO ----> 1,000h/day QA
OPS  ----> 300K+ daily
EVAL ----> write first

Utkarsh Agarwal / AI Product / Cars24 Founder's Office

AI systems doing 300K+ ops a day, built by me and run by ops.

First AI hire at Cars24. Built the team to 15, shipped 5 products to production - including AIRO with 50+ live voicebots and CLIO at 1,000+ call hours/day. Earlier: co-founded Findr — #1 on Product Hunt, 1,000+ DAUs at $70K ARR, wound down in 2024.

Self-serve voicebot/chatbot console 50+ voicebots 40% support tickets self-resolved · 2M+ minutes shipped

Ops teams author, eval, and promote bot changes without waiting on AI engineers.

Selected Work

Dossier / 2020 - 26
  • 50+ voicebots in production Cars24 / AIRO
  • 1,000h audio QA'd per day Cars24 / CLIO
  • 300K+ AI ops served daily across 5 systems
  • $70K ARR at peak Findr / 2024
  • #1 Product Hunt Day, Week, Month / 2024
  • 120+ non-tech builders trained internal hackathon
  • 15 AI engineers hired and led team built 2024-25
How I work

Write the eval first. Build the console second. Replace the model when it stops paying rent. Most AI projects die in PoC because they skip step one.

v.2026 / Bengaluru SF / Scroll
Selected work / 02 P / I / R / I

Five systems, all in production.

Each one survived contact with reality.

01

AIRO

Cars24 / Founder's Office

0->1 / Lead PM, hands on the eng 2024 - Now

Problem

Every new bot took an AI engineer a month to ship. We had a backlog of 40 use cases and a team of six. Math was not working.

What I did

I scoped, designed, and led build of a self-serve console: ops authors prompts and tools, runs them against a frozen eval set, promotes when score beats incumbent. Killed the engineer-in-the-loop entirely.

Result

50+ voicebots in production. Ops ships changes in hours, not weeks. AI team back to building harder things.

Insight

If ops cannot ship without you, you are the bottleneck. Build the console. The model is replaceable.

  • LLM orchestration
  • ASR/TTS pipeline
  • Eval harness
  • Self-serve console
  • 50+ voicebots live across 8 ops teams
  • 95-110% of human agent baseline 800-call frozen set, 3 QA graders, IRR 0.82
  • 40% support tickets self-resolved measured Q4 '25, n=1.2M tickets
  • 30 days -> 4 hrs time to ship a bot median, since console launch
02

CLIO

Cars24 / QA Intelligence

Lead PM, rubric author 2024

Problem

200+ humans listening to a sample of calls and grading them by hand. 5% sample coverage. Compliance was theatre.

What I did

I rewrote the QA rubric end-to-end with the head of ops, through 11 versions. Built an LLM grader, tuned it to within 0.4 points of senior-grader consensus. Wired it to ingest every call, every day.

Result

100% coverage. 1,000+ call hours and 8,000+ documents graded daily. The rubric became the spec - when ops disagrees with a score, they edit the rubric, not the model.

Insight

If you cannot write down what good looks like, you do not have a product. You have a wish.

  • Speech diarization
  • Rubric-graded eval
  • RAG over policy docs
  • Realtime dashboards
  • 1,000+ call hours / day ASR + diarization + grading, p95 latency 90s
  • 100% coverage, was 5% every call, every day, no sampling
  • 0.4 pts LLM grader vs. consensus 1-5 rubric, n=800 calls, 3 senior QA graders
  • +0.5% visit->token conversion controlled rollout, 6-week window, p<0.05
03

Findr

Co-founder, CTO

Co-founder, Product, Eng 2023 - 24

Problem

Your work answer is in Slack, or Drive, or Notion, or Jira. You do not know which. So you ask three colleagues instead.

What I did

Co-founded the company. Designed the product, the connectors, and the streaming-RAG architecture. Built the team to 7. Shipped to enterprise with SOC2, ISO, GDPR, HIPAA aligned, plus a self-host migration.

Result

Product of the Day, Week, and Month on Product Hunt. 1,000+ DAUs at $70K ARR. 35K+ searches in the first six months at 99% uptime. Wound down when the platforms shipped native versions in 2024.

Insight

Distribution is the moat, not retrieval. We won the channels and lost the platform shift. Both lessons cost real money.

  • Vector search
  • OAuth fan-out
  • Streaming RAG
  • Slack and Chrome native clients
  • #1 Product Hunt Day, Week, and Month - Jan 2024
  • 1,000+ daily active users no paid acquisition
  • $70K ARR at peak self-serve, no enterprise contracts
  • 35K+ searches / 6 months 99% uptime, peak post-launch
04

Chains x Variant Finder

Cars24 / Internal tooling

PM, designed the matcher 2024

Problem

Listings came in with messy variant strings. We were mismatching cars to SKUs at scale - every wrong match was a refund or a lost sale.

What I did

I designed a multi-modal matcher and a no-code workflow runner so ops could compose it into pipelines without engineering tickets.

Result

+14% used-to-listed conversion. Workflow authoring went from a JIRA ticket to a 20-minute task.

Insight

Most AI feature problems are actually data-mapping problems wearing a costume.

  • Multi-modal embedding match
  • DAG runner
  • Human-in-loop review queue
  • +14% used->listed conversion controlled rollout, 8 weeks, n=120k listings
  • Zero-code workflow authoring median author is an ops analyst, not an engineer
05

GenAI Hackathon

Cars24 / Internal program

Designed curriculum, ran the room 2024

Problem

Most of the company wanted to build with AI. None of them could. Hiring AI engineers was not going to scale.

What I did

Designed a two-day, hands-on curriculum for non-technical leads. Real APIs, real prompts, real production paths. I ran the room and pre-vetted 8 ideas as production-ready.

Result

8 of those 8 shipped. The shipped projects now run inside the company; some of their authors became permanent AI builders.

Insight

Self-serve is not a tool. It is a permission structure. Once people believe they can ship, they do.

  • Curriculum design
  • Lab notebooks
  • Production handoff playbook
  • 120+ non-tech builders trained PMs, ops leads, finance
  • 8 / 8 vetted projects shipped all running 12 months later
Background / 03 BLR mostly / SF when it matters

Six years building AI in production.

Not posting about it. Where the pager goes off and the metric has to move.

  • 2024 - Now

    Cars24, Founder's Office.

    First AI hire. Built the team to 15. Shipped AIRO, CLIO, Chains, Variant Finder. Ran the company-wide hackathon.

  • 2023 - 24

    Co-founded Findr.

    GenAI search across work tools. #1 on PH, 1,000+ DAUs at $70K ARR. Wound down when platforms shipped native.

  • Nov 2020

    WhatsApp + GPT-3, in a weekend.

    60 lines of Python on a Pi. 1,000 users by morning. Where I learned distribution beats novelty.

Work with me / 04 1 slot open / July

I take on 2-3 AI projects a quarter. If yours has to actually ship, book time.

I do not pitch. I qualify. Tell me what you are building, what is stuck, and what shipping would unlock. If I can move the needle, we will talk. If not, I will point you to someone who can.

t.me/uagarwal310 - DM ↗ calendly.com/hi-utkarsh - 30 min ↗ hi@utkarsh.cloud ↗

What I am useful for: scoping, eval design, self-serve consoles, voice/audio, RAG postmortems.

What I am not: deck reviews, idea validation, fundraising intros.