Saumya Dabi

Building AI products · Mumbai

Product manager who ships with AI agents. I take AI-native products from nothing to production:

I own the spec, the evals and the review. AI coding agents write the code.

Now
Senior Product Consultant · Alpha Realm Consulting
Before
IncrediPets · CreditQ · Overlayy · IIT Kanpur
See the work Read the latest case study
Portrait of Saumya Dabi
Saumya DabiSr Product Consultant · ARCIIT Kanpur
Platforms built 0→1 in 2026
3
Records migrated, zero integrity breaks
240K+
In product & strategy since
2023
Alpha Realm Consulting · Mumbai · since Aug 2026

About

Three years building products at startups, from marketplaces and B2B SaaS to AI-native tools. I run discovery, write the spec and the evals, then build it with AI coding agents and ship it. In 2026 alone that meant three platforms built from nothing, two of them in production.

AI Product

RAG, vector search, eval design and prompt specs, all in production.

Technical Product

Python, SQL, APIs and Lambda. I direct AI coding agents through the build and read the diff before it ships.

0 → 1

Problem, riskiest assumption, cheapest test, thinnest slice.

Data

Funnels, cohorts and denominators, named every time.

Shipped to production, measured in the field.

Selected work4 cases · 2 minis

A sample from the last three years. Client names are withheld where the engagement requires it. The numbers are not.

Organisation-management platform, zero to one

Data & evalsProduct craft
Live on AWSARC · 2026API · Mobile · Web01

A statewide field organisation ran its hierarchy on a vendor portal whose data could not be trusted. I audited it, found roughly 48K fabricated records, then specced the replacement and directed AI coding agents through a ground-up rebuild: one API, an Expo mobile app for Android and iOS, a web console and a desktop wrapper.

Records240K+migrated to the new cluster
Integrity0orphan, ancestor or area breaks
Hierarchy · root → 4 ladders → branches
  • Loaded 240K+ records with zero integrity breaks
  • Serverless on AWS Lambda behind API Gateway
  • AI assistant scoped to each user's branch, with Marathi voice in and out
Members89K+across four hierarchy ladders
Tests1,582in the automated suite
Read the full case study
DiscoveryData-integrity specCut-over plan1,582-test gate
Read the case study

AI learning platform for K-12, built on the textbook

AIData & evals
On AWS · demo dataARC · 2026RAG · LLM02

Designed for schools across Maha­rash­tra, state board and CBSE. Every answer is grounded in the textbook. Students get doubt resolution, animated AI explainers and practice questions generated live. Teachers get an overview of results where a red column is a teaching gap and a red row is a child who needs help.

Books1,046state-board and CBSE textbooks ingested
Verifier16/16seeded wrong answers caught before release
Retrieval · chunks → query → grounded answer
  • RAG over 44K+ indexed chunks on Supabase + pgvector
  • Embeddings computed locally on CPU; about $0.12 per chapter to generate
  • Privacy designed in under DPDP 2023: the model never sees a child's name or score
Demo690synthetic students across 12 schools, no real child data
Corpus44K+chunks indexed for retrieval
Read the full case study
44K+ chunk corpusGrounding spec16/16 verifier evalsMarathi voice
Read the case study

Synapse: an AI company brain that can't leak across teams

AI
In productionInternal platform · 2026Desktop · RAG03

A company of about 300 people kept its operating knowledge in Notion, Slack threads, spreadsheets and people's heads. Finance and HR material couldn't go into any shared search tool. I designed the access model and the eval set, then directed AI coding agents to build a desktop knowledge platform where people ask in plain language and get answers grounded in company documents. Department access is enforced in the retrieval layer, not in a prompt.

Users290+registered across the company
Departments7each with its own retrieval scope
Access · one query, one department's index
  • Access hierarchy shipped and shown to leadership before the AI was tuned
  • Grounded answers only: "I don't have that" instead of a fluent guess
  • Silent in-app auto-update, so weekly releases needed no IT
CadenceWeeklyreleases shipped through auto-update, to v0.4
Sign-inOTPmobile login on the existing employee directory, no new passwords
Read the full case study
Dept access modelRetrieval-layer enforcementGrounded-answer evals290+ users · 7 depts
Read the case study

A pet-care booking app, rebuilt as a marketplace

BusinessAI
IncrediPets · 2025–26Consumer marketplaceDelhi NCR + Mumbai04

Partners arrived late, and most users booked once and never came back. I led a 20+ person engineering and design team on both problems. We fixed reliability with incentives and gave people a reason to return with AI-driven features.

Retention5×growth in recurring-user rate
On-time75%partner sessions, up from under 50%
Two-sided · pet parents → hub → partners
  • Rebuilt partner incentives around a graduated penalty and rewards engine
  • AI-generated diet charts and interactive pet-training features
  • Turned a booking-only app into a marketplace
Recurring25%of users, up from about 5%
Team20+engineers and designers led
Read the full case study
Booking-only appIncentive engineAI diet + trainingMarketplace
Read the case study
CreditQ · B2B SaaS2024–25
−80%Email tickets

Clients talk to their RM directly

In-app chat and calls went straight to relationship managers. Legal and background-verification upsells lifted sales 20%+ over the year. Invoicing was automated and CRM workflows rebuilt.

Overlayy · Founding team2024
2026Acquired by Monaco

Agentic sales copilot, 0 → 1

Wrote the MVP roadmap from a teardown of 10 competing products and shaped the early go-to-market positioning. Overlayy was acquired by Benchmark-backed Monaco in Sep 2026.

Don't take my word for it. Try them.

Live demos & code2 live apps · 3 repos

JobLens first: a live job platform with JD Fit built in. Then CreditLens, and the side builds they grew out of. Every number links to how it was measured.

JobLens: every job board in one ranked feed

Job seekers check five to ten boards and a pile of alert emails, and match scores come with no reasons. I specced a three-stage ranker and directed AI coding agents to ship it as a public platform: hybrid retrieval, a deterministic 0–100 score, then a deep check of the top 10 where every requirement cites a resume line or counts as a gap. 320K+ open jobs from 13 live sources in 196 countries (Oct 2026), on $0 infrastructure. JD Fit is built in: check any job description against your resume at /check.

0.871Recall@10 on 70 known-item pairsHow it was measured 10/10fit bands correct, 0 credit without evidenceHow it was measured 0hard-filter violations in the top 20, over 70 queriesHow it was measured 719 msfeed p95 over the full indexHow it was measured
13 live sourcesBM25 + embeddings0–100 scoreCited deep check
Live app · headline projectjoblens
Try it live Check any JDRead the case study

More side builds

github.com/saumya-dabi
  1. pm-agentsMulti-agent PM copilotFour specialists: briefing, metrics, vault Q&A, PRD/ticket drafting. Tools run through a local MCP server with an allowlist. Code-level guardrails block production-DB references and destructive operations before any tool runs.Python · Claude · MCP · Jira · Notion · PostHog
  2. vault-ragLocal-first RAG over 300+ notesHeading-aware chunking into ChromaDB on-device, incremental re-indexing from a hash manifest, cited answers or "not in the notes". This is the retrieval pattern Synapse grew out of.Python · ChromaDB · ONNX MiniLM
  3. DeepQLearningReinforcement learning, from scratchQ-learning to DQN to dueling double DQN. The dueling agent solves Lunar Lander with a 252 average over the last 100 episodes; 200 counts as solved.Python · PyTorch · Gym

From strategy to shipped product.

Experience6 stops · 2019–2026

Each role pushed me one layer closer to the build.

Aug 2026 — Present

Senior Product Consultant

Alpha Realm Consulting · Mumbai
  • Owned a zero-to-one organisation platform across API, mobile and web on Lambda, from spec to agent-directed build. It holds 240K+ records with zero integrity breaks and has a 1,580+ test suite.
  • Specced and shipped a zero-to-one AI K-12 platform with RAG and pgvector, designed for schools across Maharashtra and tested end to end on a 690-student synthetic demo.
Dec 2025 — Jul 2026

Product Manager

IncrediPets · Gurugram
  • Led a 20+ person engineering and design team. Partner on-time rate went from under 50% to 75%.
  • Ran the marketplace revamp. Recurring users grew 5×.
  • Shipped Synapse, an internal AI knowledge app: 290+ users across 7 departments.
Dec 2024 — Nov 2025

Product & Operations Manager

CreditQ · Jaipur
  • Owned the B2B SaaS roadmap with a team of about 25, measured on MRR, churn and ticket volume.
  • Email tickets fell to 20% of baseline. Sales grew 20%+ over the year.
Mar 2024 — Nov 2024

Product & Strategy Manager

Overlayy · Bengaluru · Founding team
Acquired by Monaco · Sep 2026
  • Founding PM on an AI-native, seed-funded sales-intelligence copilot.
  • Owned the MVP roadmap and early go-to-market narrative.
Jul 2023 — Feb 2024

Strategy Consultant

ShowTime Consulting · Hyderabad
  • Ran ground operations end to end across 4 major and 100+ minor campaigns. Delivered research and cost analytics to leadership.
2019 — 2023

B.Tech, BSBE

IIT Kanpur
  • Head of Security, Techkriti'23. Organiser, Antaragni'21, managing a ₹50L budget across 60+ events.

Indian Institute of Technology, Kanpur

B.Tech, Biological Sciences & Bioengineering · 2019–2023

Leadership on campus

  • Head of Security, Techkriti'23Security operations across every venue of IIT Kanpur's technical festival.
  • Organiser, Antaragni'21 · Media & Publicity, Udghosh'21Led teams of 50+ and 70+ people. Grew digital reach 75% YoY to 40M through a 5,000+ student ambassador programme.

The build step changed. The sequence didn't.

How I work6 stages

Being able to build working software with AI agents changes what the cheapest test is. It doesn't change the order, and a prototype that answered a question still isn't a spec.

Stage 1

Problem

Written down, with a named user.

Stage 2

Riskiest assumption

The one that kills it if wrong.

Stage 3

Cheapest test

Often a throwaway prototype an agent builds in an afternoon.

Stage 4

Thinnest slice

The smallest credible version, shipped.

Stage 5

Instrument

Every metric states its denominator.

Stage 6

Decide

Keep, kill or cut deeper.

Product depth, technical fluency.

Product
Product strategy0→1 discoveryPRDs & specsRoadmappingA/B experimentationFunnel & cohort analysisOKRsAgile / Scrum
AI / Data
LLM applicationsRAGVector searchEval designPrompt specsPythonSQLPostHogMongoDB AtlasSupabase · pgvector
Build & tools
AI coding agents (Claude Code)AWS LambdaAPI GatewayREST APIsGitFigmaJiraNotionCRM automation

Thinking out loud, most days.

Writing36 essays

Essays on AI product management, evals, agents and the craft of shipping 0→1.

Latest · Fintech & SaaS

Deactivation is a product decision, not a billing task

In B2B SaaS, involuntary churn often stems from rigid billing logic and blunt deactivation scripts. Treat account cutting as a core product journey.

Nov 3, 2026 · 6 min readRead the essay
India Tech

Hyperlocal growth stalls on supply coverage before user demand

Oct 31, 2026 · 5 minTeardown

A model lineup is a pricing portfolio

Oct 27, 2026 · 5 minAI Product

Human overrides should expire

Oct 24, 2026 · 4 min
All writing36 essays

Let's build the next

0→1
Email me Download résumé Mumbai · IST · replies by email
LinkedInin/saumya-dabi GitHubsaumya-dabi
Copied