A clever model can survive a bad answer. A factory cannot shrug off a missed exception, and a fleet does not care that the prompt was ambiguous. Somewhere between the smooth chat window and the untidy physical world, artificial intelligence acquires consequences. Anirudh Pupneja has been moving toward that border for most of his brief public career.
His route looks, at first, like a quick tour of the AI boom: computer science at BITS Pilani, natural-language research at IIIT Delhi, generative-AI infrastructure at Coinbase, then Y Combinator with Redapto. Look closer and a steadier pattern appears. Each stop asks a version of the same unfashionably practical question. How does a machine earn trust when its output must be measured, corrected and used?
Redapto now describes itself as an agentic AI platform for the critical operations behind manufacturing, logistics, energy and defense. Its software is meant to join live signals, specialized agents and human operators in one governed system. The language is industrial: sites, fleets, suppliers, uptime, exceptions. It is about as far from a blinking chat cursor as software marketing can travel.
First, teach the machine to grade
In 2023, Pupneja appeared among ten authors of H-AES: Towards Automated Essay Scoring for Hindi. The setting was MIDAS Lab at IIIT Delhi. The problem came with an awkward absence: automated essay scoring had a long history in English, while Hindi lacked a dedicated corpus for the task.
The team translated an established English essay dataset, assembled a smaller real-world Hindi corpus and compared several approaches. Classical feature-based models sat beside LSTMs and fine-tuned transformers. The work was published in the AAAI proceedings, and its dataset, rubric and code were released publicly.
This was research about evaluation before evaluation became the favorite word of every AI product meeting. An essay score is a compact output, but producing it responsibly means confronting language, scarce data, prompt behavior and agreement with human judgment. The machine must do more than sound plausible. Somebody has to decide what good looks like.
“I believe, when it comes to AI, our experimentation phase is hardening into a production phase.”Anirudh Pupneja, December 2025
Pupneja’s student record supplies a smaller, more tactile scene. During a BITS Pilani practice-school project, he worked on a classification algorithm and features for an Android application. In his published reflection, he wrote about learning TensorFlow, app development, Java and literature review while crossing between computer science and another technical domain. The vocabulary is charmingly pre-founder: unfamiliar tools, a helpful mentor, valuable work. No grand theory required.
His GitHub account preserves the workshop floor from that period. There is a Hindi-English transliteration tool built for MIDAS Lab, an essay-scoring repository, a saliency-based visual-attention model, an Android client for a distributed network and a full-stack food-ordering system for the BITS Goa campus. Fifteen public repositories make an untidy and useful counterweight to the polished founder biography. They show a student trying Java, Kotlin, Python and JavaScript against different kinds of problem. The projects do not pretend that a career arrives as a single clean thesis. They show the thesis being assembled, one repo at a time: collect a signal, build an interface, make the moving parts communicate. Even the campus takeaway application belongs in the story. Long before Pupneja was talking about agents coordinating suppliers and workflows, he had code coordinating hungry students and food orders. Scale changes. The inconvenience of systems talking to one another is remarkably loyal.
Then, make it run inside a company
After BITS, Pupneja built generative-AI platform and internal agent infrastructure at Coinbase. Public biographies keep the account concise. A later Redapto launch profile says some of those automations save more than $2.5 million a year. The exact machinery stays offstage, but the change in environment is clear. A research model can end with a table of results. An internal platform has users, permissions, failure modes and a Monday morning.
It also gave Pupneja a production-scale version of the problem he had already encountered in research. Models need context. Outputs need a standard. Corrections need somewhere to go. The organization has to remember what the model forgets.
systems · documents · teams
plan · coordinate · act
policy · exceptions · review
That loop would become Redapto’s organizing idea. Pupneja founded the company in 2025 with Cheril Shah, whom he had first met while they were fine-tuning NLP models in the same research lab, before large language models occupied every investor memo. Shah is also a co-author on the Hindi essay-scoring paper. The founding connection came with a shared working artifact, not merely a shared enthusiasm.
A startup changes its nouns
Redapto entered Y Combinator’s Fall 2025 batch in San Francisco. Its first widely circulated product story centered on customer support. The company checked interactions across chat, voice and email, looked for quality and knowledge gaps, and used what it found to improve training material, procedures and knowledge bases. The pitch was a closed feedback loop: observe the work, find the error, repair the system that produced it.
By August 2026, the company’s public language had moved into the physical world. The current site talks about procurement, production, logistics, field service and work that cannot stop. It promises a governed operating layer that can connect fragmented systems without replacing the tools a team already uses. Operators define policy, review exceptions and retain command.
A pivot is often narrated like a personality transplant. Redapto’s public record reads more like an expansion of the original sentence. Customer support supplied conversations to inspect and procedures to update. Physical operations supply sensor signals, documents, suppliers, machines and field decisions. The nouns changed. The loop grew teeth.
Publish research comparing ways to score Hindi essays when training data is scarce.
Build generative-AI and internal agent systems at Coinbase.
Found Redapto and join Y Combinator’s Fall batch.
Position agents inside governed, physical-world operations.
The orchestrator’s wager
Pupneja set out his broader view in a December 2025 essay about AI’s “production era.” He wrote about open-source models, self-improving agents, human feedback and what he called an orchestrator era. His most useful observation concerned verification. Reinforcement learning needs domains with scoring functions and trajectories that can be corrected mathematically. Without a reliable way to verify an output, humans remain the experiment.
That concern becomes sharper around a production line or an energy system. Model quality matters, but latency, policy and exceptions can matter just as much. A capable agent with vague authority is a fresh species of colleague: tireless, literal and alarmingly confident. The sensible response is not to admire it from a demo. Give it boundaries, logs and a manager.
Redapto’s design language reflects this. “Human in the loop” is specified as policy, review and command rather than offered as a ceremonial reassurance. The software is supposed to act continuously where it has permission and bring the right exceptions to people. It is an argument for delegated autonomy, with the delegation written into the product.
“For now, we are the orchestrators. But for how long?”Anirudh Pupneja
There is a pleasing tension in that question. Pupneja is building systems meant to orchestrate work while describing humans as the present orchestrators of AI. The roles may trade pieces over time. His published estimate is patient: translating verifiability across domains could take a decade and require techniques beyond current reinforcement learning.
Patience is useful here. Agent companies are often sold with verbs borrowed from magic. Redapto uses the plainer grammar of operations: connect, evaluate, coordinate, review. Its founder’s public path gives those verbs some history. He began with a machine asked to grade an essay. He moved to machines asked to help a company run. Now he is asking what they may safely do when the work reaches warehouses, fleets and factory floors.
The work after the answer
The chat window made AI legible to millions of people. It also made the technology look complete at the moment a sentence appeared. Pupneja’s work starts one beat later. Where does that answer travel? Which system can it change? Who approves the unusual case? How will the next decision improve?
Those questions are less theatrical than a new model release. They are also where software becomes part of an institution. Redapto’s wager is that critical industries want intelligence woven into the operation, with enough context to be useful and enough governance to be trusted.
A Hindi essay, an internal Coinbase workflow and a logistics exception share no obvious surface. Underneath, each requires a standard, a feedback loop and a human who knows when the system has missed the point. Pupneja has spent years approaching that same structure from larger and more consequential angles. The screen was the easy part. Now comes everything it touches.