Clinical trials
Runs the recruitment pipeline for a trial, so enrolment keeps moving whether or not anyone at the site has time to chase it this week.
Building intelligence people can use.
Arhat Labs is an applied AI lab conducting fundamental research and building AI-native operating systems: systems that hold how a particular kind of work actually runs, do the repetitive parts, and hand back the decisions that need a person.
ResearchAgent infrastructureProductsReal-world impact
Research and deployment belong together.
We believe the most important advances in AI will emerge from a continuous loop between research and real-world use. Our research helps us understand and govern increasingly capable systems. Our products reveal the practical problems that research must solve.
Research
Understanding how AI systems reason, represent information, and act.
Business agents
Building agents that can understand workflows, use tools, coordinate information, and complete meaningful work.
Operating systems
Putting that infrastructure to work as the system a team runs its day on, in one domain at a time.
Enterprise workflows, one domain at a time.
Each of these came out of one workflow we set out to understand properly. The practice is the same every time; the domain is what changes.
Industrial manufacturing
Holds a plant's production, quality and maintenance picture in one place, and turns each shift's decisions into a record at the moment they are made rather than at the end of the week.
Childcare
Carries the administrative weight of a nursery — enrolment, ratios, records, families — so it stops being carried by the people who are supposed to be in the rooms.
Commercial real estate
Answers the research question behind a commercial building once, and keeps the answer current as ownership, tenancy and timing move underneath it.
Building operations
Keeps a building's requests, work orders, vendors and inspection schedules moving as one operation, rather than as whatever the manager on duty happens to be holding.
Hiring and talent evaluation
Makes interviews comparable, so the decision at the end rests on what candidates actually demonstrated rather than on who remembers which conversation.
Small and midsized business
Gives a company with no AI team the thing an AI team would build first: a system that knows how the business actually runs, and can act on that.
Retail and bespoke commerce
Closes the gap between a bespoke order and the supplier purchase it implies, without either end of it being retyped by hand.
From complex workflow to intelligent system.
We build production systems, not demonstrations. That difference shows up in what happens after the first working version.
Understand
We study the real workflow, including exceptions, incentives, existing tools, and human judgment.
Build
We create the data, integration, retrieval, agent, and interface layers required to operate reliably.
Evaluate
We measure outputs against real examples and capture structured human feedback.
Deploy
We introduce the system gradually, with clear permissions, auditability, and human review.
Improve
The system becomes more useful as it accumulates feedback, operational context, and institutional knowledge.
A demonstration only has to work once, for an audience that wants it to. A production system has to work on the exception, on the malformed document, on the Friday afternoon nobody is watching — and it has to be reviewable when it does not.
Researching the foundations of trustworthy agency.
Two threads run through our research: what a system is actually doing internally, and what it should be permitted to do in the world.
arXiv preprint, June 2026
Behavioral Governance for Autonomous AI Agents: The AgentBound Framework
Research into runtime governance for autonomous agents. AgentBound evaluates proposed agent actions against delegated authority, behavioural policies, and environmental constraints before execution.
Active direction
Mechanistic interpretability
Research into how neural networks represent concepts, process information, and produce behaviour. The goal is to make advanced AI systems more understandable and easier to evaluate.
This is ongoing work. Nothing in this direction has been published or peer-reviewed — we will link results here when there are results to link.
AI that works inside the business, not beside it.
Most organizations do not need another chatbot. They need systems that can understand their information, coordinate existing software, preserve institutional memory, and perform work under appropriate human supervision.
Capabilities
Information
- Document intake and structured extraction
- Research and knowledge retrieval
- Integration with existing tools and data systems
Action
- Workflow orchestration
- Email and communication agents
- Voice agents
- CRM and pipeline automation
Governance
- Evaluation and quality monitoring
- Human approval workflows
- Audit logs and cost visibility
- Secure deployment within enterprise environments
Where this work tends to matter most
Healthcare and clinical research. Financial services. Commercial real estate. Recruiting. Professional services. Operationally complex small businesses. These are industries where the workflows are genuinely complex, the information is unstructured, and the consequences of getting something wrong are real.
Built by experienced engineers and entrepreneurs.
Arhat Labs was started by two people who have spent their careers building systems that had to work — at scale, for real users, with real consequences.
Pranay Gupta
Co-founder
Pranay is a former Staff Engineer and engineering leader at Google with more than 15 years of experience building internet-scale systems. He led teams across Google Ads, Search, and Cloud AI, and launched multiple generative AI products for large enterprises.
Prakhar Gupta
Co-founder
Prakhar is a two-time healthtech founder and former AI product leader. He launched one of the first AI voice agents deployed in health insurance — work covered by The Wall Street Journal and Forbes.
Let's build something genuinely useful.
We work with organizations that have valuable workflows, complex information, and a serious reason to apply AI.