Teaching Enterprise AI to Ask Permission: How Vishnu Vardhan Reddy Yeruva Is Building AI That America’s Regulated Industries Can Trust
The SAP architect and patent-pending inventor is tackling the problem that keeps generative AI out of pharmaceutical, medical device and manufacturing operations: proving that a system only used what it was allowed to see.
By Will Jones, Geek Insider
Most conversations about generative AI start with the model. Which one is smartest, which one is fastest, which one tops the latest benchmark. Vishnu Vardhan Reddy Yeruva starts somewhere else entirely.
“I got into this because the interesting problem was never whether an AI system could answer the question,” he says. “It was whether it was allowed to.”
Yeruva is an SAP Business Technology Platform (BTP) Solution Architect and full stack engineer at MyGo Consulting Inc., an SAP Gold Partner and Inc. 5000 company headquartered in the United States. His work sits at a junction that most of the AI industry has been slow to reach: getting retrieval-augmented AI, multi-agent copilots and intelligent automation to run inside regulated enterprises, where every document is governed by access controls, data residency rules and audit obligations. He is a named inventor on a pending U.S. patent application in enterprise retrieval optimization, and two of his research papers have been accepted at IEEE conferences in 2026.
The Governance Gap Holding Back Enterprise AI
American companies have adopted generative AI at remarkable speed. The governance around it has not kept pace, and nowhere is that gap more costly than in the industries that make medicines, medical devices and the physical goods the economy runs on.
“In a regulated company, an AI system that gives the right answer out of a document the user was never permitted to open has not succeeded,” Yeruva says. “It has created an incident.”
The problem is structural. Most retrieval tooling was designed for open or uniformly readable collections of documents. Inside a pharmaceutical or medical device company, the document estate is anything but uniform. What a given employee may see depends on their role, their site, the clinical study or product line they support, and the regulatory jurisdiction they work under. Conventional AI pipelines treat those permissions as an afterthought, filtering results after the fact. The outcome is a system that either leaks restricted content or is restricted so heavily that it stops being useful.
Yeruva identifies three open problems in the field:
- Permission-aware retrieval. Retrieval systems that do not natively respect the access-control model of the data beneath them.
- Grounded output. No reliable way to verify that a model’s structured output is actually drawn from the source document rather than merely plausible.
- Auditability. Regulated industries must be able to reconstruct why a system produced a given answer, and most AI tooling was never built to support that.
“In a validated environment, a system that is usually right but cannot show its work is not deployable,” he explains.
The stakes reach well beyond any single company. America’s AI Action Plan, released by the White House in July 2025, calls for accelerating AI adoption across American industry and emphasizes that reliable AI requires rigorous and standardized evaluation, including national standards for AI in healthcare. Pharmaceutical, medical device and advanced manufacturing are among the sectors where the United States most needs AI to deliver, and among the sectors where compliance requirements make deployment hardest. Engineering that closes the governance gap is what allows U.S. firms to use AI where they are most constrained.
From Mathematics to Enterprise AI
Yeruva’s path into technology began with a love of mathematics. He started his professional software engineering career in India in 2019, then moved to the United States to pursue graduate study, earning a Master of Science in Software Engineering from San Jose State University in 2023.
Two turning points shaped the work he does today. The first came in his graduate coursework in statistics and machine learning, which moved his focus from general application development toward applied AI. The second came after he joined MyGo Consulting in August 2023, when he was assigned to client engagements in regulated industries. There, compliance constraints on enterprise AI stopped being an abstract topic and became the day-to-day design problem.
“Build things that have to survive a compliance review,” he advises engineers entering the field. “Constraints are where the real engineering problems are.”
Since joining MyGo, Yeruva has progressed to SAP BTP Solution Architect, owning the technical design of the platform and AI systems he delivers. He has served as architect and lead engineer on enterprise engagements for Fortune 500 pharmaceutical, medical device and specialty manufacturing companies, completing eight projects and training six engineers along the way. He also contributes to MyGo’s SAP BTP Center of Excellence and was recognized internally as Employee of the Month.
His technical credentials span the full stack of enterprise AI delivery. He holds multiple SAP certifications covering backend development, AI development and data analytics on SAP BTP, along with Fiori, Integration Suite and solution architecture. That breadth matters: as Yeruva puts it, each skill is common on its own, but the combination of SAP platform architecture, full stack delivery and applied AI retrieval design under regulated-industry constraints is much less so. Engineers who know AI are readily available. Engineers who also know how to get an AI system through a pharmaceutical company’s validation and compliance review take years of engagement-level exposure to develop.
Building AI That Knows What It May See
The centerpiece of Yeruva’s work is what he calls authorization-aware enterprise retrieval. Rather than bolting permission checks onto the end of an AI pipeline, his approach enforces the permission model of the source systems as part of retrieval itself, so that the answers a system produces can be reviewed and audited by compliance teams. He is a named inventor on a pending U.S. patent application covering his work in enterprise retrieval optimization. Because the application is pending, he declines to discuss its technical mechanism, describing only the problem it addresses and the outcome it enables.
The beneficiaries, he says, are regulated enterprises and their employees, who gain usable internal AI tooling without taking on compliance exposure. Downstream, patients and customers benefit from faster operational and maintenance decisions.
That retrieval work sits alongside a portfolio of production systems Yeruva has designed and delivered for U.S. pharmaceutical, medical device and manufacturing enterprises:
- Multi-agent operational copilots that support enterprise operations teams inside regulated environments.
- Automated inventory and EDI reconciliation, replacing manual cross-system matching with continuous automated reconciliation.
- Generative-AI predictive maintenance for manufacturing assets.
- Enterprise customer data platforms built on SAP BTP.
Across these engagements, Yeruva reports that the systems he delivered cut operational processing times by roughly 50 percent and improved system performance by about 30 percent.
Case Study: Reconciling a Fortune 100 Supply Chain
His largest project illustrates the approach. For a Fortune 100 pharmaceutical and medical device manufacturer, inventory positions were tracked across SAP ERP, third-party logistics providers and trading partners exchanging EDI documents, and those sources disagreed constantly. Reconciliation was largely manual. Teams pulled reports from separate systems, matched records by hand and investigated discrepancies after the fact.
“In a regulated supply chain, an unreconciled discrepancy is not just an accounting problem,” Yeruva notes. “It is a traceability problem.”
He designed and built an automated reconciliation layer on SAP BTP that ingests EDI transaction sets alongside ERP inventory movements, normalizes them into a common representation, matches them automatically and surfaces only genuine exceptions for human review. The system handles the partial-match and timing-difference cases that cause simple matching to fail, so the exception queue reflects real discrepancies rather than noise. The result was a shift from periodic manual reconciliation to continuous automated reconciliation, with discrepancies caught in near real time instead of at month end.
Beyond the Enterprise: Research and Sahayi AI
Yeruva’s contributions extend into the research community. Two of his papers have been accepted at IEEE conferences in 2026, ICCNCT and ICOSICS. He is also the creator of Sahayi AI, an iOS application exploring generative AI at the mobile edge, which was the subject of a February 2026 interview with TechBullion.
What Comes Next, and Why It Matters Beyond One Company
Yeruva’s agenda for the next several years builds directly on the problems he has already worked on. He is continuing his work on authorization-aware retrieval, and he is conducting an independent study on grounded faithfulness verification for enterprise document extraction: checking, field by field, that every value an AI system extracts from a document can be traced to a specific span of the source text. The study uses only public, permissively licensed datasets (CORD, SROIE and Kleister) and open-source tooling, so that its methods and results can be shared openly with the wider research community. He also plans to extend his governance-first approach from pharmaceutical and medical device work into adjacent regulated sectors.
“I would like the default assumption in enterprise AI to be that retrieval respects permissions and that every extracted field can be traced back to the line it came from,” he says. “Right now both are optional.”
The significance of that goal is national rather than local. Yeruva’s work is not tied to a single product or a single employer. As a consulting architect, he carries the same methods across multiple U.S. enterprises, and the open problems he is working on, permission-aware retrieval, grounded extraction and auditability, are shared by every regulated company trying to deploy AI. Patent filings in the United States add to domestic intellectual property, while published research and open evaluation methods let other engineers and organizations build on his results.
For American workers and consumers, the benefits are concrete: faster and more reliable operations at U.S. pharmaceutical and manufacturing employers, and fewer manual errors in the supply chain and maintenance processes that affect whether products reach patients and customers on time and to standard.
Yeruva sees the coming years as a turning point for the field. “Enterprises adopted generative AI much faster than they built the governance around it,” he says. “The next few years of enterprise AI will be spent closing that gap, and that work is less glamorous and more consequential than the model releases.”
If he succeeds, the result will be an enterprise AI landscape where American companies can put their own document estates to work without accepting access-control risk. For the industries that keep the country healthy and productive, that may matter more than the next benchmark record.