AI Product Builder — turning frontier models into products people trust.
Zero to one, shipped to hundreds of millions.
Varshine builds products at platform shifts.
Founding PM of Microsoft’s first frontier reasoning agent, Researcher.
Now redefining how products get built in the AI era — across Enterprise and Consumer tech.
The Work
01
Researcher Agent.
Microsoft 365 Copilot.
Dynamics 365 Customer Service.
Twilio.
From enterprise data quality to a frontier reasoning agent to everyday work surfaces like Copilot — shipped at Microsoft scale.
02
Microsoft Copilot — Researcher Agent: 0→1 MVP to production in 5 months.
Model Council: hack week to production.
New products stood up from scratch — Microsoft’s first agentic AI system, grown to millions of users. Built the playbook — behavior specs, quality bars, launch rigor — where none existed.
03
Researcher: Microsoft’s first frontier reasoning agent.
Model Council: multi-model deliberation, in production.
System prompts as product surface. Evals as the steering wheel. I write the production prompts, build the LLM-as-judge frameworks that score them, and make ship calls on evidence — not vibes.
Work
Nine years building products; six building 0-to-1 AI, agentic, ML, and data products. Hands-on builder at the intersection of product, UX, research, engineering, safety, and GTM. Founding PM on Microsoft’s first agentic AI system — now one of the most-used deep research products in the world. Before that: AI for customer service, platform products, and enterprise solutions.
Researcher · Founding PM
Founded and launched Researcher 0→1 in Microsoft Copilot — a long-horizon agent for multi-step deep research across web and enterprise data, built with OpenAI Research. Own product strategy, UX, core quality, and AI safety, including the first joint Microsoft × OpenAI Deployment Safety Board review. Now one of the most-used enterprise deep research agents in the world.
Multi-Model Systems · AI Builder
Lead Researcher’s multi-model strategy: launched model choice by bringing Claude into Researcher, shipped the Critique capability, and envisioned and built Model Council — multiple frontier models reason over the same task; a judge model analyzes their work and synthesizes the final answer. I prototyped it as a hack project, then shipped it to production — authoring the judge’s system prompt myself and tuning it against deterministic and LLM-as-judge metrics that define what a high-quality response tastes like. Pioneered PM-led prompt tuning at scale: ideation → prototype → prompt iteration → automated LLM-as-judge + human scoring → production readiness.
Quality, Evals & Post-Training
Built and operationalized the agent-quality flywheel — eval sets, rubric graders, golden outputs, and calibrated LLM-as-judge scoring — and scaled the expert human-evaluation program into a post-training data pipeline. Shipped Microsoft’s first in-house post-trained deep research model, deployed across public and Gov clouds.
Beyond the product
01
Drove the Gates Foundation’s adoption of Researcher for vaccine research, operating hands-on as a forward-deployed engineer.
02
Launched the Claude model family in Microsoft Copilot as part of leading Microsoft’s multi-model strategy.
03
Model orchestration for intelligent auto-routing across model families by user task, while optimizing compute.
Customer Service Copilot · Support AI/ML
Led productization and growth of Dynamics 365 Customer Service Copilot — conversational AI for enterprise support built on early GPT models. Ideated and launched AI ticket summarization, case timeline highlights, and inline email assist within case management. Built ML-based automated ticket creation across ~50M annual support cases.
Segment Data Platform
Built integrations with Google, Salesforce, Meta, and Adobe, and developed the Integrations Contributors program opening Segment’s platform to partner teams. Launched Multi-Instance Destinations — a cross-team effort spanning 30+ services and one of the platform’s largest launches.
Windows Developer Technologies
Advised and architected solutions for 25 large enterprises across Asia — 400+ hours of technical consultation. Translated recurring user needs into product capabilities and presented brownbag sessions at 30+ customer sites on web development, virtualization, and Azure.
Education
01
Master of Science, Technology Management · 2019–2020
02
Bachelor of Technology, Computer Science · 2012–2016
Method — how I ship agents
A kolam is drawn fresh each morning: dots first, then a single continuous line that loops through every one without lifting. Good agent development is the same shape — the loop only works if it never breaks.
01
Build the smallest real thing. Working demos beat decks — hack week is a legitimate ship vehicle.
02
The system prompt is product surface. Behavior specs live in the prompt, and I author the ones we ship.
03
LLM-as-judge plus calibrated human panels. Taste becomes rubrics; rubrics become graders.
04
Behavior spec, safety review, launch call. Frontier capability earns trust through rigor.
05
Findings become the next roadmap bet — and the line returns to the first dot.
If quality can’t be scored, it can’t be improved. I turn “this feels better” into datasets, assertions, and rubric graders a team can climb.
The words a model reasons with shape everything users feel. I write production system prompts myself — and defend them with evals, not opinions.
I led Microsoft’s first joint deployment safety review with OpenAI. Trust is what lets frontier capability reach millions of people at work.
Personal
Weekend builds and hackathon projects — the same loop, off the clock.
Project · 02 — OpenClaw Hackathon
A personal AI assistant built at the OpenClaw hackathon. Full demo below.
Project · 01 — Multi-Model
The Model Council pattern as a standalone web app — multiple models deliberate on a prompt, a judge synthesizes the answer.
Try it ↗Hosted service currently suspended
Project · 03 — Agents
An experimental agent build — code and write-up on GitHub.
GitHub ↗Project · 04 — Memory
A self-improving memory harness for LLMs — code and write-up on GitHub.
GitHub ↗Plus a running stack of weekend hacks — local model fine-tuning, hardware and robotics with GPT models, and more.
A note on depth
Frontier agent quality, model orchestration, and safety work is confidential by nature. The real depth is in the conversations — happy to go there.
Currently
Leading product for frontier AI agents at Microsoft AI — Researcher, multi-model strategy, and agent quality in Microsoft Copilot. Writing the prompts, building the evals, making the ship calls.
Open to conversations about frontier AI product leadership, speaking, and collaborations in AI × product.