AI product manager — evals, agentic systems, data platforms.
I own the measurement layer: the evals, guardrails and routing that decide whether an AI system is allowed to ship.
Today at EY's AI Center of Excellence, putting GenAI into production inside client environments across BPO, aerospace, healthcare and insurance. Before that, Nielsen's audio data platform.
work
Roles, selected deployments, evaluation practice and capabilities.
Product Manager, Enterprise AI Platform
EY, AI Center of Excellence · Oct 2025 — Present
I own EY's AI platform for data transformation, GenAI insights and agentic decisioning, and the eval and review layer that every client deployment must pass before rollout.
- 1,000+ — case human-labelled eval set every release is scored against
- 10% → 90% — conversation QA coverage at Everise, 50,000+ interactions a day
- 92% — automated scores matching human auditors
- Own the eval and guardrail layer. Every release is scored offline against a 1,000+ case human-labelled set, 5% of production traffic is sampled for human review, and PII is redacted before anything reaches a model.
- Scaled conversation QA for Everise, the BPO behind Apple, Microsoft and Airbus support, from a 10% manual sample to 90% automated auditing across 50,000+ interactions a day. Automated scores match human auditors 92% of the time. Defined the mistake taxonomy and scoring rubric with the client's QA leads, ran auditor calibration on disagreements, and owned the reviewer workflow. Delivered inside the client's locked-down VDI.
- Shipped a conversational agent for an insurance carrier that runs live database queries mid-conversation to give personalised answers. Owned how tool calls and failures are logged and surfaced for review. Client reported first-contact resolution up 28% and wait times down 60%.
Product Manager, AI Data Platform
Nielsen Media · Apr 2024 — Oct 2025
I owned the roadmap for Nielsen's audio data platform, and wrote a fair amount of it. The work was equal parts product and infrastructure: what the data promised downstream, and whether the pipeline could keep that promise.
- 70% — reduction in data lag
- 99.9%+ — uptime at P99
- 15% — lower platform cost
- Shipped the embedded analytics that closed multi-year enterprise contracts on a $650M audio measurement line, which grew 5% year over year.
- Designed and wrote the Python microservices behind ETL orchestration on an S3 data lake and a 12-node EMR-Spark pipeline handling multi-terabyte days.
- Negotiated data SLAs with data science and operations: P95 lag under five minutes, above 99% accuracy.
Product Manager, B2B Subscription
Eblity · Jun 2022 — Mar 2024
Customer discovery, pricing and the growth roadmap for a B2B subscription sold into institutions, with the long sales cycles that come with them.
- 35% — revenue growth year over year
- 250% — ROI on institutional partnerships
- 45% — faster setup
- A/B tested signup and landing pages; client acquisition up 20% month over month.
- Worked alongside sales and marketing through institutional procurement cycles.
Associate Product Manager, Consumer Platform
Sparklin · Nov 2020 — Jan 2022
My first product role, and where the habit started: ship, measure, and let the experiment settle the argument.
- 30% — user acquisition, month over month
- 15% — engagement, month over month
Automated QA at 50,000 interactions a day
SELECTED DEPLOYMENT · Everise — BPO for Apple, Microsoft, Airbus
Everise audited roughly a tenth of its interactions by hand. We took auditing to 90% coverage across more than 50,000 interactions a day, running entirely inside their locked-down VDI environment.
- 90% — of interactions audited, up from 10%
- 92% — scoring parity with human auditors
- 50,000+ — interactions a day
Parity with human auditors was the number that made the system acceptable. Without it, coverage would have been a statistic nobody trusted.
Agentic case resolution for repair and overhaul
SELECTED DEPLOYMENT · Honeywell Aerospace
Scoped and shipped a case-resolution agent with Honeywell's service team, with tool calling and a human in the loop wherever the model was not confident.
- 83% → 97% — CSAT
- 93% — contractual threshold
- Renewed — contract outcome
In-call decision support
SELECTED DEPLOYMENT · Firstsource — healthcare BPM
Decision support in front of agents mid-call, so they could resolve queries faster and move into higher-value work rather than simply closing tickets.
- ~$0.02 → ~$0.50 — revenue per call
A conversational agent on live policy data
SELECTED DEPLOYMENT · Insurance carrier
Most conversational agents answer from a static corpus. This one runs live database queries during the conversation, so the answer reflects the policy the customer actually holds.
- +28% — first-contact resolution, client reported
- −60% — wait times
The bar: nothing ships without proof it works
EVALUATION PRACTICE
Enterprise AI does not fail because the model is weak. It fails because nobody can show, in terms a risk function accepts, that the system behaves. The eval layer is the argument you hand to the people who can stop your launch.
That reframes the work. Evals are not a QA chore at the end of a roadmap; they are the artefact that gets AI past security, risk and procurement.
Offline scoring against a labelled set
EVALUATION PRACTICE
Every change is scored against a human-labelled set of more than a thousand cases. It is slow to build and unglamorous to maintain, and it is the only thing that turns a subjective 'this feels better' into a number two teams can argue about.
Human review on sampled production traffic
EVALUATION PRACTICE
Five percent of production traffic is sampled for human review. Offline sets describe the world you imagined; sampling describes the one you got. When the two disagree, the labelled set is what needs updating.
Guardrails and PII redaction
EVALUATION PRACTICE
Redaction and guardrails are what make deployment possible inside client environments — locked-down VDI, no customer data crossing the network boundary. This is the part client security and risk teams read line by line.
Model routing under a cost ceiling
EVALUATION PRACTICE
Routing cut inference cost by 40%. The constraint that mattered was not the saving; it was proving quality did not move while the saving happened — which is only checkable if the eval layer already exists.
Parity as the acceptance criterion
EVALUATION PRACTICE
When AI replaces human judgement, coverage is not the headline number — agreement is. At Everise, 92% scoring parity with human auditors was the threshold that let automated QA stand in for people.
AI and evals
CAPABILITIES
- Eval design and offline scoring
- Guardrails and PII redaction
- Agentic workflows with tool calling and human-in-the-loop
- RAG and semantic layer design
- Model routing and LLM cost control
Data and platform
CAPABILITIES
- Python, SQL, Spark
- AWS — S3, EMR, Athena
- Microservices and ETL pipelines
- Semantic data modelling and real-time data systems
- Embedded analytics and self-serve BI
Product craft
CAPABILITIES
- 0→1 delivery in regulated environments
- Roadmap ownership and pricing
- Go-to-market and A/B testing
- SLA negotiation
- Direct work with client security, risk and operations teams
Education
FOUNDATION · 2016 — 2020
RCC Institute of Information Technology — B.Tech, Electronics & Communication.
Top startup, TiECON Kolkata.
writing
Essays and product notes.
Aug 2026
Unified infrastructure for collection, processing, storage and distribution — and why the enterprise now depends on it.
Read on Substack
Apr 2026
Weekly AI updates: actionable tips, top startups and state-of-the-art breakthroughs in ten minutes.
Read on Substack
Apr 2026
Weekly AI updates: actionable tips, networking opportunities and breakthroughs worth your attention.
Read on Substack
Apr 2026
The first edition of the weekly roundup.
Read on Substack
Dec 2025
Moving beyond vibes to build a rigorous grading system for the un-measurable.
Read on Substack
May 2025
Long term, long term, long term.
Read on Substack
May 2025
A two-minute read on where AI actually helps a product manager.
Read on Substack
Apr 2025
Must-follow resources for product builders.
Read on Substack
Jan 2025
A teardown of Uber's mass-premium subscription and what it assumes about price sensitivity.
Read on Substack
Jun 2022
What the Lindy effect is, why it matters, and how it affects the things you build.
Read on Substack
Jun 2022
A product concept: lightweight plans for the group chat where plans already happen.
Read on Substack
Mar 2022
Product note on zero-cost assistance for surgeries.
Read on Substack
Mar 2022
Product note on building a habit of investing in gold, a rupee at a time.
Read on Substack
Mar 2022
Product note on a decentralized platform connecting artists directly with fans.
Read on Substack
Mar 2022
Product note on making a team's video content searchable and useful.
Read on Substack
Mar 2022
Product note on writing and scheduling for people who publish in public.
Read on Substack
