Dhaval Khamar
Product leader · Agentic AI · Data & BI · Audiences & Measurement

Agentic AI built on data people trust.

For 20+ years I've turned complex workflows into software. Today I lead a team of six PMs building AI agents, audiences and measurement for programmatic advertising, on data platforms and BI people can rely on.

See selected work Get in touch
signal_from_noise · fit(model)R² 0.00
raw data insight decision
12
AI, audience and measurement launches in 2025–26
$2B+
In ads billing run through Poseidon Analytics, the platform I led
+30%
BI adoption after launching Looker MCP
$32M
A year cut from the AWS bill
Built atYahooCapgeminiOracle
01 · About

Engineer, consultant, product leader

I'm a product leader who started as an engineer and has spent 20+ years turning complex professional workflows into software. My work sits where three things meet: agentic AI, the data platforms underneath it, and the business intelligence people use to make decisions.

I lead the Audiences, Measurement & AI product line at Yahoo, where my teams ship LLM-powered search, a natural-language assistant and agent interfaces built on MCP and agent-to-agent protocols. I also led Poseidon Analytics, the billing, warehouse and BI platform behind $2B+ in ads billing, and earlier ran an enterprise BI and machine-learning roadmap for finance.

Before product, I spent 14 years in consulting at Oracle and Capgemini, leading M&A integration, large programs and process consulting for enterprise clients.

My domain depth is programmatic advertising: DSP workflows, audience targeting, CTV and retail media measurement, and multi-touch attribution. I find the real problem with customers and engineers, write the strategy and specs, size each bet with a P&L model, and stay hands-on through launch.

02 · How I lead

Four principles I build by

01

Data before agents

An agent is only as dependable as the data it acts on. I sequence the work accordingly: data foundations, then speed, then AI-driven discovery. AI built on data nobody trusts doesn't get used.

02

Check the plumbing first

An audit found only 13K of 394K conversion rules were active. Before debating strategy or models, I verify the data a product runs on actually does what everyone assumes.

03

Every bet gets a P&L

No major initiative enters the quarterly plan without a P&L model, so revenue and advertiser-value trade-offs are explicit before engineering starts.

04

Guardrails are product features

Agents get blocklist checks, bounded outputs, automatic labeling and human approval from version one. Privacy-safe patterns like k-anonymity and compute-to-data are designed once and reused.

03 · Focus areas

Three capabilities, two domains

dim_01 / Agentic AI

AI Agents & LLM Products

Natural-language copilots, semantic search, AI-written metadata at scale and troubleshooting agents, plus the interfaces that let agents use a product directly: MCP servers and agent-to-agent marketplaces. Guardrails and human approval are designed in.

  • AI agents
  • MCP · A2A
  • Copilots
  • Semantic search
  • LLM evaluation
dim_02 / Data platform

Data Platforms

The foundation agents and analytics depend on: billing and data warehouses, ads data pipelines, privacy-safe clean rooms, purchase signals, and cloud cost discipline that took a $72M yearly AWS bill to $40M.

  • Data warehouse
  • Pipelines & streaming
  • Clean rooms
  • Data governance
  • Cloud cost
dim_03 / BI

Business Intelligence

One trusted version of the numbers: consolidating four analytics systems into one, defining clear metrics, and putting AI on BI with a Looker MCP server so people can ask questions without opening a dashboard.

  • Looker
  • AI on BI
  • Metric design
  • Finance BI & ML
  • Measurement
domain_01 / Audiences

Audiences

Making audience data easy to find, build and activate, for people and for agents: a natural-language copilot, semantic search across about 4.8M segments, faster activation and better-scaled segments.

24–48h → 4–6h custom audience activation+153% scale on undersized segments12K+ AI searches in 14 weeks
domain_02 / Measurement

Measurement

Outcome measurement advertisers can trust: privacy-safe clean-room CTV measurement, conversion attribution integrity, retail media closed-loop reporting and purchase signals for bidding.

61 advertisers on Netflix CTV measurement at launch46 advertisers' mis-credited conversions fixed71h → 5h attribution root-cause analysis
04 · Experience

From engineer to product leader

role · company
200020052010201520202026
Product Management Lead, Audiences, Measurement & AIYahoo DSP · San Francisco
2021 – now
Principal Product Manager, Enterprise BIYahoo
2017 – 21
Program Manager · Principal Architect · M&A integration leadCapgemini · Chicago, Kansas City, Bangalore
2007 – 2017
Senior Consultant, Oracle FinancialsOracle · Johannesburg, South Africa
2003 – 07
Senior ProgrammerApplitech Solutions
2000 – 03
05 · Selected work

From problem to impact

Agentic AI · Audiences

Audience Copilot & Smart Audience Search

A natural-language assistant, semantic search and an MCP tool so people and agents can find and build audiences.

Problem

Ad traders could only find audiences by exact name in a catalog of about 4.8 million segments.

Approach

Search over LLM-generated descriptions, backed by a backfill that raised description coverage from about 15% to every new and existing audience.

12K+

searches from 327 advertiser seats in 14 weeks, 0.57s median response

Agentic AI · Measurement

AI troubleshooting agent

An agent for the conversion-attribution pipeline.

Problem

An audit found only 13K of 394K conversion rules were active, and finding the root cause of a data issue took days.

Approach

Fixed mis-credited conversions for 46 advertisers, set a next-day, 99.5%-accuracy target and introduced an AI agent for diagnosis.

71h→5h

root-cause analysis time

BI · AI on BI

Poseidon Analytics & AI on BI

A billing, data warehouse and BI platform, opened up to natural-language questions.

Problem

Separating the ads business from consumer systems meant standing up separate billing and analytics, with reporting spread across four disparate systems.

Approach

As product manager, I led Poseidon Analytics, which generates $2B+ in ads billing and partner payment reporting. We consolidated four analytics systems into it, deduplicated data and defined clear metrics, then launched an MCP server on its Looker instance in May 2026 so people can query data without logging in to Looker.

4→1

analytics systems, $1M+ in savings, and 30% higher BI adoption with Looker MCP

BI · ML

Enterprise BI & ML roadmap

Finance analytics and predictive support.

Problem

Yahoo's finance organization needed one roadmap for BI and machine learning that delivered measurable savings.

Approach

Built and delivered the finance BI and machine-learning roadmap, including ML-based predictive analytics. Delivered $1.25M in cost savings.

+200%

BI usage, with 60% fewer support tickets

Data platform

AWS cost optimization

Bringing cloud cost in line with revenue.

Problem

Running a large ad platform on AWS cost $72M a year, a direct hit to the platform's margin.

Approach

Two lenses. Technical: mapped each product to business components, technology components and AWS services, then cut data retention left behind by years of growth (about 20% saved). Business: bid-volume analysis showed 20–30% of lower-funnel bid requests could be dropped with negligible impact (about 15% more).

$72M→
$40M

yearly AWS cost, about $32M saved each year

Measurement · Data platform

Netflix CTV clean-room measurement

Privacy-safe attribution for connected-TV ads.

Problem

Advertisers couldn't measure Netflix CTV campaigns without moving user-level data.

Approach

Clean-room attribution using k-anonymity and compute-to-data, built as a pattern now being reused for Disney, Walmart and Spotify.

61

advertisers at GA, Sep 2026

06 · M&A & Programs

Before product: M&A, programs and process

Before moving into product, I spent 14 years in consulting at Oracle and Capgemini, working across three continents: Bangalore, Johannesburg, Tokyo, Kansas City and Chicago. That's where I learned to run large programs, untangle business processes and deliver against a hard number.

M&A integration · Capgemini

$140M

Annual savings from leading the technology side of the Motorola Wireless spin-off and its integration with Zebra Technologies, a $4B deal. Delivered on schedule and 20% under budget, against a $105M savings target.

Program management · Capgemini

$3M → $10M

Grew a key client account to $10M a year as Program Manager, leading delivery teams of up to 30 across onshore and offshore locations.

Solution architecture · Capgemini

$10M

Revenue from a project-billing product I designed for engineering firms, adopted by Black & Veatch, AECOM and Burns & McDonnell. It won the iGATE Innovator of the Year award in 2014.

Process consulting · Oracle

Oracle Financials

Collections, receivables, payables and expenses implementations for Cell C and Mutual & Federal, plus data migration and a customer loyalty program for UK Co-operatives.

07 · Writing

Notes from the field

  • Substack · Nov 2026
    Retail media · Measurement

    Closed Loop Isn't the Same as True

    Retail media fixed attribution's data problem, not its counterfactual problem. Why closed-loop ROAS overstates some tactics, how ROAS-optimizing bidders drift toward shoppers who would have bought anyway, and what networks should change.

  • Substack · Sep 2026
    AI adoption

    The Model Isn't the Bottleneck. Your Data's Reputation Is

    An essay on AI adoption: AI projects stall when people stop trusting the data underneath them, and that trust erodes much faster than it builds.

  • Yahoo · Apr 2026
    Agentic AI · Data

    Why Data Is the Queen in the Agentic AI Era

    Co-written with Srinivas Bhagavatula. An AI agent is only as dependable as the data underneath it, so accurate, governed, real-time data matters more than an impressive demo, with identity resolution, open-standard integrations and human approval controls as the example.

Subscribe on Substack

08 · Toolkit

What I work with

AI products

  • LLM applications
  • AI agents, MCP & A2A
  • Semantic search
  • Guardrail design

Audiences & measurement

  • Programmatic (DSP)
  • Audience targeting
  • CTV & retail media
  • Multi-touch attribution

Data & platforms

  • Pipelines & data lakes
  • Clean rooms
  • Streaming (Flink)
  • AWS

Consulting & delivery

  • M&A integration
  • Program management
  • Process consulting
  • Client account growth

Product craft

  • Customer discovery
  • Strategy memos & PRDs
  • P&L and scenario models
  • Coaching PM teams

Education & certifications

  • B.E., Electronics & Communications · Dharmsinh Desai University
  • Mini-MBA, General Management · Loyola University Chicago
  • Deep Learning Algorithms and Applications · UC Santa Cruz
  • Product Management Certificate · General Assembly  ·  Certified Scrum Product Owner (CSPO) · Scrum Alliance
09 · Contact

Let's talk about what your data can prove.

I'm open to senior product leadership roles, advisory work and speaking. Topics I'm glad to dig into:

  • Retail media incrementality
  • AI agents in ad platforms (MCP, A2A)
  • Clean-room CTV measurement
  • Data strategy before AI
  • Building PM teams for data products
Email[email protected] LinkedInlinkedin.com/in/dhavalkhamar Substackdhavalkhamar.substack.com LocationSan Francisco, CA Résuméavailable on request