Data Products | ex-AWS

Tushar Sharma

I ship AI-native products teams actually adopt. Semantic layers, agent tools, and the instrumentation that proves what works.

Currently exploring senior Product, Data Product, AI Product, and Strategy & Ops roles.

Tushar Sharma
San Francisco Bay Area ttsharma018@gmail.com LinkedIn
Supio
10,000+
end users on the reporting I owned
In-product dashboards and QBR reporting for 500+ law firms, on a unified case model spanning four case-management systems.
AWS
$82M
unmanaged revenue surfaced
AWS customer-mapping dashboard exposed spend that had no account owner.
Supio
<5 min
from question to insight
AI-native semantic layer at Supio replaced 24-hour analyst Q&A across Sales, CS, RevOps, Ops, and Product.
Supio
$32M
annual revenue supported
Invoicing went from a manual Finance process to an in-product feature I product-managed, cut over at zero reconciliation variance.

About

I'm a data and product operator with eight years across AWS, startups, and consulting. As the founding data hire at Supio I built the analytics function from scratch, then shipped the AI tooling that runs on top of it, including a governed semantic layer that 150 people across five functions query for themselves. I've owned measurement, attribution, and the operating cadence behind board reporting. The through-line is knowing which problems need a model, which need a process, and which need a UX fix.

How I Operate

Instrument before you model.

If the metric that matters isn't tracked, that's the first product gap to close. Every good decision I've shipped started life as a measurement problem.

When UX is the bottleneck, don't add a model.

My biggest north-star win was a navigation change with zero model changes, proven with a randomised A/B rather than argued. AI judgment includes knowing when AI isn't the answer.

Self-serve beats service.

Queues don't scale; systems do. The best analytics product removes the analyst from the loop, including when the analyst is me.

In immature categories, build thin on open source.

Committing your metric layer to an unstandardized vendor category is the most expensive lock-in. Keep the definitions yours; swap the serving layer when the market matures.

Case Studies

Case study

AI-Native Semantic Layer

Operational data was locked behind a 24-hour analyst queue; new reports took days. I designed an AI-native semantic layer end to end: dbt marts plus YAML specs on DuckDB, served through an MCP server and queried from Claude Desktop. The design decision that mattered was a tiered query path, where governed metrics answer first, a read-only fallback restricted to curated marts discloses its own provenance, and novel questions route through human validation. 150 people across Sales, Customer Success, RevOps, Operations, and Product built their own reports against it. Insight time collapsed from 24 hours to under five minutes, and two analyst hires per business function were avoided.

AI-Native Architecture Semantic Layer Self-Serve
Read the case study →
Time from question to insight
BI + analyst queue 100
Semantic layer + Claude 2
Case study

PM Intelligence Tool

PMs had no direct line to customer intelligence scattered across Gong, HubSpot, Slack, Notion, and Intercom. I designed a multi-agent system on the Claude API that surfaces it: an analyst, a scriber, and a strategist over a shared skill library. Early runs flagged churn risk that wasn't there, so I added the evidence rules that made it trustworthy, where every insight cites its sources, single-source findings stay suppressed, and a pain point is checked against the roadmap before it counts as a gap. Product used it weekly for PRDs and business reviews. Customer-research time fell roughly 60% and PRD drafting went from hours to minutes.

0→1 Product AI Agents Adoption
Read the case study →
Customer-research time
Before 100
After 40
Case study

Instrumentation, then Experimentation

Engagement metrics were everywhere, but nothing measured whether customers ever found the features we shipped. The loop I run: find what's missing, instrument it, diagnose friction via user research, ship a scoped experiment. Release popovers only fired on login and users kept sessions open for days, so I replaced them with a persistent release tab and tested it as a firm-level cluster-randomised A/B across 300 firms. Week-one feature activation went from 30.0% to 47.4%, a 58% relative lift that held in every firm-size stratum. It became the company's north star.

Instrumentation A/B Experimentation Decision Quality
Read the case study →
North-star metric
Baseline 100
After Marquee 158
Case study

AI Operations Measurement

Vendor annotators reviewed and corrected AI output over medical and claims records, and none of that work was measured. I instrumented it down to per-task effort, then built the edit-rate metric tracking how often a human had to correct the model. A rising edit rate meant degradation, which put the question in front of engineering with evidence attached. Edit rate on critical events fell from 60% to 25%, and knowing where annotator effort actually sat is what told us which steps were worth automating.

AI Quality Measurement Instrumentation Model Monitoring
Read the case study →
Edit rate on critical events, percent
Before instrumentation 60
After 25
Case study

Revenue Routing Strategy

One Salesforce customer can own many AWS accounts, and legacy routing had scattered a startup's accounts across four other segments, so Startup teams developed accounts other segments were paid for. Three orgs each had their own number and the argument was about credit. I anchored the definition in the funnel, required program evidence on top of identity matching so it couldn't over-attribute, corrected my own org's routing gap before asking anyone else to move, and took it to SVP sign-off in a 6-pager. $82M of revenue moved into the segment already managing it.

Identity Resolution Influence Without Authority Metric Definition
Read the case study →
Revenue moved to the right segment, $M
Before the standard 0
Year to date 82
Case study

Attainment Reporting Migration

Four functions at AWS each had their own attainment reporting, and the views didn't agree, so leadership reviews opened by arguing about which number was right. I consolidated the estate onto QuickSight across four to six teams, validating every rebuilt report against the original and standardizing definitions on the way. It gave back up to 840 resource hours a year. It also taught me the more useful lesson, which is that leadership sign-off and actual adoption are different measurements and only one of them is the goal.

Cross-Team Program Adoption BI Migration
Read the case study →
Resource hours returned a year
Before consolidation 0
Up to 840

More Work

Customer-Facing Reporting

Law firms ran on four different case-management systems, so the same case looked different in each one, and reporting to them lived in spreadsheets. I interviewed firm executives about what they wanted in front of them, designed a unified case data model that reconciled all four into one backend, and shipped in-product dashboards covering case outcomes and product adoption. 500+ customers and 10,000+ end users came off spreadsheets, and the QBR reporting built on it went into renewal conversations.

Customer-Facing Analytics Data Modeling Discovery

Trial Conversion & Expansion

Supio ran free and paid trials and nobody could say what a converting one looked like. I instrumented feature adoption into the warehouse and found the signals that separated trials that closed from trials that stalled: depth of use, breadth beyond the champion, and speed to first value. Then I acted on them, working with the VP of Implementation, SVP of Engineering, and Head of Product to remove the UX friction and carry unmet requirements into the roadmap. Trial conversion rose 20% and expansion revenue 15%.

PLG Activation Analysis Conversion

Pricing, Packaging & Tier Design

Partnered with Strategy and Revenue Operations on pricing, packaging, and tier design, supplying the usage and unit-economics analysis the options were weighed against. Also gave Operations leadership the per-task cost and productivity data behind sizing staffing against upcoming workload. Strategy and RevOps owned the decision; I owned the numbers it was made on.

Unit Economics Pricing Strategy Partnering

Marketing Attribution Rebuild

Roughly $7M in annual program spend ran against pipeline nobody could trace. I connected spend to CRM in the warehouse, diagnosed 30% of pipeline as unattributed, and rebuilt the UTM taxonomy and lead-source capture behind it. Multi-touch attribution then showed last-touch over-crediting bottom-of-funnel channels. Marketing reallocated, and blended CAC fell 15% over two quarters.

Attribution Marketing Measurement Data Contracts

Investment Qualification Platform

AWS needed a defensible way to choose which startups to back with credits. I owned the scoring framework behind that call, integrating PitchBook, Crunchbase, CBInsights, and Dealroom to score funding, GTM signals, and market fit. Scoring and segmentation models raised program efficiency 30%.

Product Strategy GTM Signals Analytics

Billing & Revenue Infrastructure

Revenue recognition ran on manual, error-prone effort. I architected interim pipelines on HubSpot, AWS Glue, and Mode that cut manual invoicing work 75% while the real feature was built. We piloted a third-party metering platform, found it couldn't model our custom rates and multi-year ramps, and I carried the build-versus-buy recommendation to a decision. I then designed the cutover reconciliation, a full-base comparison run as parallel billing cycles until variance was zero, so no invoice cycle was missed. The system supports $32M in annual revenue.

Revenue Ops Vendor Management Product Management

Semantic Layer & Self-Serve Analytics

Four functions reported from their own numbers, and the contested definitions were the ones that mattered most. I designed the entity model and the governance behind 40+ standardized KPIs across Salesforce, billing, and external data, led four analysts through the build, and defined the row- and role-level access rules the platform team implemented. Named owners, written change proposals, and versioned definitions with lineage kept it trusted. 550 people worked from it and ad-hoc reporting fell 35%.

Self-Serve KPI Strategy Enablement

Campaign Efficiency Measurement

Three organizations at AWS measured campaign performance three different ways. I owned the Redshift KPI layer over Finance spend, program data, and Adobe Analytics, and designed the conformed program dimension and identity bridge behind full-funnel measurement from first web hit to closed deal. It became the source of record for recurring business reviews across five orgs.

Full-Funnel Measurement Identity Resolution Alignment

Startup360 Analytics Platform

A customer-intelligence and seller-productivity product on Redshift, S3, and QuickSight that became the most-used seller tool in the AWS BI suite. I ran it for more than a year; it reached 5,000+ stakeholders and lifted seller productivity 16% year over year, with governed self-serve datasets published from it.

0→1 Product Self-Serve Seller Productivity

Experience

Supio

Sep 2024 – Aug 2026

Data Products

Founding data hire at an early-stage AI SaaS company. Built the analytics function, self-serve tooling, and measurement frameworks from scratch, then used them to shape the product roadmap.

  • Designed and shipped an AI-native agentic analytics platform, with dbt marts and a governed semantic layer exposed to Claude through an MCP server. Governed metrics answer first and novel questions route through human-in-the-loop validation. Time-to-insight fell from a 24-hour SLA to under 5 minutes and two analyst hires per function were avoided.
  • Shipped a multi-agent PM tool on the Claude API over Gong, HubSpot, Slack, Notion, and Intercom, with evidence rules requiring cited sources and corroboration before a signal surfaces. Product used it weekly for PRDs and business reviews. Customer-research time fell roughly 60% and PRD drafting went from hours to minutes.
  • Connected roughly $7M in annual program spend to CRM pipeline, diagnosed 30% of pipeline as unattributed, and rebuilt the UTM taxonomy and multi-touch attribution model behind it. Blended CAC fell 15% over the following two quarters.
  • Owned measurement for the human-in-the-loop annotation pipeline, instrumenting operational signals to per-task effort and cost, and building integrity checks that flagged overstated vendor volumes. Workflow automations cut processing time 30%.
  • Built the interim billing and revenue pipelines on HubSpot, AWS Glue, and Mode, cutting manual invoicing effort 75%, then piloted a third-party metering platform, found it couldn't model our custom rates and multi-year ramps, and drove the decision to build invoicing in-product. Designed the cutover reconciliation that ran parallel billing cycles to zero variance. The system supports $32M in annual recurring revenue.
  • Designed the Fullstory and PostHog instrumentation for feature discovery, activation, and engagement, then ran a four-week firm-level cluster-randomised A/B across 300 firms that lifted week-one feature activation from 30.0% to 47.4%, a 58% relative lift, doubled repeat use, and became the company north star.

Amazon Web Services

Jan 2021 – Jun 2024

Analytics & Insights Products

Promoted to lead product analytics for the AWS startup program, owning data products that guided credit, discount, and GTM decisions.

  • Built a startup-scoring platform on PitchBook, Crunchbase, CB Insights, and Dealroom whose scoring and segmentation models qualified startups for credit and discount programs and raised program efficiency 30%.
  • Designed and governed a centralized semantic layer standardizing 40+ KPIs across Salesforce, billing, and external data, leading four analysts through the build and running the sign-off process that unified contested definitions into a single trusted source. 550 people worked from it and ad-hoc reporting fell 35%.
  • Owned campaign efficiency reporting for AWS Startups, designing the conformed program dimension and identity bridge that enabled full-funnel measurement from first web hit to closed deal.
  • Consolidated the SMGS organization's attainment reporting from Tableau to QuickSight across four to six teams, validating every rebuilt report against the original and driving the legacy suite to deprecation. Gave back up to 840 resource hours a year.
  • Designed the seller-alerting pipelines on Redshift, S3, and Airflow that detected account signals and delivered each alert with a proposed next best action. Contract renewals rose 20% and customer outreach 35%.
  • Ran Startup360 on Redshift, S3, and QuickSight for over a year, the most-used seller tool in the AWS BI suite, reaching 5,000+ stakeholders and lifting seller productivity 16% year over year.

Amazon Web Services

May 2020 – Jan 2021

Business Intelligence

Joined AWS supporting Revenue and Sales Operations with data pipelines, dashboards, and go-to-market analytics.

  • Resolved a three-way definition dispute over how misrouted revenue should be measured, drove alignment through a 6-pager and SVP sign-off, then shipped the revenue-routing strategy on Redshift, Python, and Tableau that surfaced $82M in unmanaged revenue.

Perficient Inc.

Aug 2019 – Apr 2020

Data Engineering

Built and optimized healthcare data pipelines and dimensional models in a compliance-driven consulting environment.

  • Redesigned healthcare claims models with Kimball dimensional techniques, cutting processing errors 25% and securing compliance approval from key stakeholders.
  • Optimized SQL, Python, and shell pipelines, improving operational efficiency 20% and supporting a 15% improvement in clinical decision-making.

ADL Group

Sep 2016 – Jul 2017

Business Intelligence

Early analytics and BI role spanning SQL optimization and dashboarding for a real-estate data business.

  • Optimized T-SQL stored procedures and queries, improving query performance 30%.
  • Built Tableau dashboards tracking building-occupancy and KPI trends that helped drive 20% customer-acquisition growth.

Skills

Product & Measurement

Product roadmapping & discovery PRDs & 6-pagers KPI & metric definition A/B testing & experimentation Funnel, cohort & retention analysis Stakeholder alignment Build vs. buy & vendor management Operating cadence (WBR/QBR) Multi-touch attribution UTM taxonomy & lead-source capture

AI

Claude Code Claude API agents in production MCP server development Agentic workflow automation Prompt & context engineering Conversational analytics Human-in-the-loop validation design AI quality measurement in production Agent behaviour tracing (Laminar, ClickHouse) Cursor

Data Engineering

SQL Python PySpark Shell scripting (Unix, Bash) dbt Airflow AWS Glue ETL/ELT Dimensional modeling (Kimball) Data contracts & validation Monitoring & reconciliation Anomaly detection RLS & RBAC Git & PR review

Warehouses, Cloud & BI

Redshift BigQuery Databricks DuckDB PostgreSQL MySQL SQL Server (T-SQL) Snowflake AWS GCP Azure Tableau Looker QuickSight Power BI Mode Hex Sigma Excel

Instrumentation & GTM Systems

Fullstory PostHog Google Analytics Adobe Analytics Google Ads HubSpot Salesforce Label Studio Airtable

Education

Syracuse University

M.S., Information Management · GPA 3.79 / 4.0

2017 – 2019

University of Mumbai

B.E., Computer Engineering · GPA 3.75 / 4.0

2013 – 2016