Talks on the reality of enterprise AI.
Invited speaker at Snowflake, Adobe and NVIDIA conferences. Every topic is something I ran, built or governed at a Fortune 5.
Six talks.
Keynotes, panels and executive briefings.
Growing an AI organisation from 4 to 57+
What had to change at each step, from hiring to standards to eight-figure vendor negotiation.
Full abstractClose
More than seven years at UnitedHealth Group took me from leading four engineers to 57+, a multi-million dollar budget with full P&L responsibility, and a seat alongside business stakeholders and executive sponsors. The talk covers what had to change at each step, from hiring to architectural standards to eight-figure vendor negotiation, and why the engineering never stopped for the leadership to stay credible.
AI governance that does not stop delivery
Review criteria engineers accept rather than route around.
Full abstractClose
I founded and lead the AI/ML Review Board at UnitedHealth Group, which sets enterprise architectural standards and Responsible AI guidelines, and I sit on it as lead R&D judge. The talk covers model risk management, review criteria that engineers accept rather than route around, and how that governance sits alongside $37.2M+ in combined annual savings delivered by the same organisation.
Healthcare AI at the scale of the enterprise
117,808 decisions a day inside HIPAA, SOC 2, ISO 27001 and GDPR.
Full abstractClose
117,808 automated prior-authorization decisions a day, 100K+ claims a month adjudicated in under a minute instead of ten, and 10M+ daily API transactions across 10PB-scale infrastructure, all inside HIPAA, SOC 2 Type II, ISO 27001 and GDPR. The talk covers synthetic data, digital twins and PHI/PII encryption as the things that make experimentation possible rather than the things that block it.
Agents in production, not in demos
What breaks when agents touch real enterprise systems, and what survived.
Full abstractClose
Snowboard is a native desktop agentic RPA platform I architected end to end to replace an eight-figure vendor contract: 3,200 bots in production at 99.7% uptime, serving roughly 410 users and removing multiple millions of dollars a year of RPA spend. The talk covers the low-level Windows automation hooks, computer vision for UI detection and process mining underneath it, and what actually breaks when agents touch real enterprise systems.
Secure execution for the code your models write
Five layers between an agent and the infrastructure, adopted by 15 teams.
Full abstractClose
CAGE is 16,000+ lines of Rust that put five layers of defense in depth between an LLM agent and the infrastructure, across nine languages. The layers are application validation, orchestrator isolation, container namespacing, gVisor kernel interception and host cgroups. It cut cybersecurity threats by 78% and TCO by 40% and is now used by 15 teams, and the talk walks through the threat model, the architecture and the adoption path.
Small specialised models beat big general ones
A 6.75M-parameter model deciding 43 million authorizations a year.
Full abstractClose
MNM is a 6.75M-parameter numeric transformer deciding 43 million prior authorizations a year. Trained on PolicyDAG synthetic causal data, it reaches 77.17% accuracy against an 81.9% Bayes ceiling on that data, 94.2% of what the data allows, after training on a consumer RTX 2080 in 48.6 minutes for $0.02. The talk covers column tokenization, PolicyDAG synthetic causal data and multi-task checkpoint supervision, and where a compact model is the right answer instead of a frontier one.
Bios, ready to copy.
Three lengths, all current.
Short bio
69 words · For a programme listing or a chair's introduction.
Show the bioHide the bio
Yash Sharma leads an AI organisation of 57+ engineers and researchers at UnitedHealth Group (Fortune 5), with a multi-million dollar budget and full P&L responsibility, and founded the AI/ML Review Board that governs enterprise AI there. Research engineer by trade, he holds 213 U.S. patents licensed by five of the biggest names in big tech and deep tech, has 700+ publications, and is a 6× consecutive AI World Champion.
Medium bio
147 words · For an event page or a panel description.
Show the bioHide the bio
Yash Sharma leads an AI organisation of 57+ engineers and researchers at UnitedHealth Group (Fortune 5), with a multi-million dollar budget and full P&L responsibility, working directly with business stakeholders and executive sponsors. He founded and leads the AI/ML Review Board that governs enterprise AI, negotiated eight-figure vendor contracts, and is a founding technical partner of Optum & UHG Ventures. He also architected the flagship systems his organisation runs, each from a proof of concept he built himself: Snowboard, an agentic RPA platform running 3,200 bots at 99.7% uptime that replaced an eight-figure vendor contract, CAGE, a 16,000-line Rust sandbox for code written by LLMs, and MNM, a 6.75M-parameter transformer deciding 43 million prior authorizations a year. Together the work has delivered $37.2M+ in combined annual savings. Research engineer by trade, he holds 213 U.S. patents and 700+ publications, and is a 6× consecutive AI World Champion.
Long bio
268 words · For a keynote page, a press release or a moderator's notes.
Show the bioHide the bio
Yash Sharma leads an AI organisation of 57+ engineers and researchers at UnitedHealth Group (Fortune 5), 32 direct and 25 indirect, organised as forward deployed pods across 80+ enterprise products. He works directly with business stakeholders and executive sponsors, carries a multi-million dollar budget with full P&L responsibility, founded and leads the AI/ML Review Board that sets the enterprise's architectural standards and Responsible AI guidelines, negotiated eight-figure vendor contracts, and is a founding technical partner of Optum & UHG Ventures. Over 7+ years he progressed from a software engineering analyst leading four engineers to the organisation he runs today.
He architected all six flagship systems that organisation runs, each as a working proof of concept, and took them to enterprise scale with his teams. Snowboard, a native desktop agentic RPA platform, runs 3,200 bots at 99.7% uptime and replaced an eight-figure vendor contract. CAGE, a 16,000-line Rust sandbox, gives LLM agents five layers of defense in depth and cut cybersecurity threats by 78% and TCO by 40% across 15 teams. MNM, a 6.75M-parameter numeric transformer, decides 43 million prior authorizations a year after training on a consumer RTX 2080 for $0.02. CRYO, a multi-agent framework, was adopted by every core line of business. Together they account for $37.2M+ in combined annual savings.
Research engineer by trade, he holds 213 U.S. patents licensed by five of the biggest names in big tech and deep tech, has 700+ publications, and is a 6× consecutive AI World Champion. He began in computer vision research at DRDO's Centre for AI and Robotics, where his object-tracking networks were adopted by four defense agencies.
Headshot.
Cleared for event listings, programmes and press. Credit Yash Sharma. Do not crop the face.
- Format
- PNG, full-length portrait, transparent background
- Dimensions
- 950 × 1448 px
Facts for introductions.
Six lines for a chair with thirty seconds.
- Scope
- Leads an AI organisation of 57+ at UnitedHealth Group (Fortune 5), 32 direct and 25 indirect, with a multi-million dollar budget and full P&L
- Governance
- Founded and leads the AI/ML Review Board, and is founding technical partner of Optum & UHG Ventures
- Delivered
- $37.2M+ combined annual savings, 3,200 bots at 99.7% uptime, 43M+ prior authorizations a year
- Patents
- 213 U.S. patents, licensed by five of the biggest names in big tech and deep tech
- Publications
- 700+, first author on the majority
- Competition
- 6× consecutive AI World Champion at the Dell, Apple & NVIDIA AI Championships
Invite me to speak.
Conferences, panels, executive briefings and internal AI summits. I reply within 48 hours.