Skip to content
Yash Sharma

I lead enterprise AI at a Fortune 5. I still build.

I lead 57+ engineers and researchers building enterprise AI at UnitedHealth Group.

Book a conversation
Read the writingLinkedIn
Yash Sharma, standing, in a dark suit and white shirt
I'm Yash Sharma. Research engineer by trade. I build Fortune Five enterprise AI from scratch and lead the teams that run it. I have 213 United States patents, licensed by five of the biggest names in big tech and deep tech. More than 700 research publications. Six time consecutive AI World Champion. I run the organization, not just the architecture. I lead 57 engineers, with full profit and loss responsibility for the budget. And I've negotiated eight-figure vendor contracts. Across the enterprise, my work delivers multi-million dollar savings every single year. Six systems, live inside the enterprise. Snowboard, CAGE, MNM, CRYO, FinTwinOS and the GPU Center World Model. From research, through production, to the boardroom. I set the AI strategy, design the architecture, get the systems into production, and govern how they run. I'm open to executive leadership, board and advisory roles. Book a conversation.
engineers and researchers led
57+
engineers and researchers led
enterprise products in production
80+
enterprise products in production
combined annual savings
$37.2M+
combined annual savings
U.S. patents
213
U.S. patents

What I own.

Organisation
57+ engineers and researchers
Economics
Multi-million dollar budget, full P&L
Enterprise AI
80+ products in production
Outcomes
$37.2M+ combined annual savings
Governance
AI/ML Review Board, founded and led

Six flagships. The latest of 80+ products in production.

I architected the initial versions of all six and scaled them with my teams.

Six systems. One operator. Here they are. Snowboard: my automation platform, built from scratch. 3,200 bots, 99.7 percent uptime, multi-million dollar savings a year. I built CAGE: more than sixteen thousand lines of Rust. Five defense layers around AI-written code, and 78 percent fewer threats. Then MNM: a 6.75 million parameter Transformer, deciding more than 43 million prior authorizations a year. I built CRYO, the multi-agent framework every line of business builds on. 60 percent more productive. FinTwinOS is my live twin of a financial institution. Agents rehearse decisions before they're real. My GPU Center World Model rehearses where each job lands, and says how sure it is about every option. I lead, not just build. 57 engineers, full profit and loss responsibility. I work directly with executive sponsors. Book a conversation.

Snowboard. The automation platform that replaced a licensed robotic process automation estate.

Bots in production
3,200
Bots in production
Uptime
99.7%
Uptime
Dollar annual savings
Multi-million
Dollar annual savings

Explore

CAGE. Secure execution for AI-generated code, entirely inside the enterprise perimeter.

Reduction in cybersecurity threats
78%
Reduction in cybersecurity threats
Lower total cost of ownership
40%
Lower total cost of ownership
Teams adopted enterprise wide
15
Teams adopted enterprise wide

Explore

MNM. A 6.75M-parameter numeric transformer deciding prior authorizations at national scale.

Prior authorizations a year
43M+
Prior authorizations a year
Decisions a day
117,808
Decisions a day
Dollar annual savings
Multi-million
Dollar annual savings

Explore

CRYO. The multi-agent framework every line of business now builds its automation on.

Dollar annual savings
Multi-million
Dollar annual savings
Team productivity increase
60%
Team productivity increase
Core lines of business adopted
All
Core lines of business adopted

Explore

FinTwinOS. A live twin of a financial institution where AI rehearses decisions before they are real.

Explore

GPU-Center World Model. Counterfactual admission control for GPU fleets, with calibrated uncertainty on every option.

Explore

All systems and outcomes

The engineering never stopped. The scope grew.

How I lead, in full

  1. Build when ownership compounds.

    Licensing, lock-in, security and cost decide build versus buy.

  2. Bound the risk. Then ship.

    Put the security boundary where the regulator needs it and let the teams move behind it.

  3. Start with the economics.

    Cost per decision, workflow and adoption before the choice of model.

  4. Central standards, local execution.

    One review board sets standards and release gates. Forward deployed pods do the shipping.

The depth beneath the flagships.

Every model I build in a regulated healthcare enterprise has to answer four questions. What does one decision cost? Can it read the whole history? Does data stay inside our boundary? Can a regulator audit it? A single frontier model call answers none of them well. Too slow, too costly, hard to audit. So I built a family instead. Each one fits its question, and each is small enough to run on our own hardware. The volume was always there. The architectures weren't. The numeric decision transformer I built sits at the center. It reads each case as columns, not words. Six point seven five million parameters, deciding more than forty three million prior authorizations a year. Eight more architectures surround it. One reads a member's whole history in one pass. Another sends each claim to the experts that fit it. One spends longer thinking only on hard cases. Another reads messy interchange files byte by byte. One drafts a whole letter within constraints it can't leave. Another learns a single view of the patient. One adapts to a new plan as it runs. Another keeps its memory compressed for agents that run all day. A fraction of the cost per decision. The whole history for context. Data on premises. One model serves many programs. Fewer models to govern. Nine architectures I built for healthcare, live in the enterprise. Book a conversation.

Specialised models for enterprise healthcare.

9 architectures built for cost, auditability, long context and regulated data.

Explore the research

Enterprise healthcare digital twin.

Rehearse policy, capacity and AI decisions before they touch real operations.

Explore the twin

Selected thinking.

All 86 essays

Let’s talk.

Executive roles, board and advisory work, venture diligence, speaking. I reply within 48 hours.

Book a conversationconnect@yashsharma.ai