The first machine-learning system attached to Ishan Amlekar's name had an unusually democratic instinct: let the human argue back. A model offered a prediction. A person adjusted it. The correction could become new training data, and the machine could return a little wiser. The work began during his 2019 internship at supply-chain software company Kinaxis, while he was still an undergraduate. Years later, the resulting family of interactive machine-learning inventions would produce three United States patents. By then, Amlekar had moved from correcting predictions to teaching software how to navigate whole sequences of work.
This is the useful thread through an otherwise itinerant early career. He moved through insurance data science at Intact, education software at Syngli, software engineering at Qualcomm and behaviour simulation at autonomous-vehicle company Nuro. The sectors changed with the regularity of university co-op terms. The recurring subject was the exchange between a person and a system: what the user means, what the software observes, and what either side can learn from the other.
At Scribe, where Amlekar became a founding machine learning engineer in June 2023, the exchange became a dataset. Scribe records a user's actions as they complete a digital task and turns that activity into step-by-step documentation. Those traces contain something more valuable than screenshots. In sequence, they describe work: where it begins, what context matters, and which action ought to follow.
The most revealing line in Amlekar's career may be the shortest: prediction, correction, retraining.A pattern first formalized in the Kinaxis patent work
01 / Early signalsA club with 500 members and no shortage of jobs to do
Long before San Francisco, there was the Network of Digital Enthusiasts at Delhi Public School Gurgaon. Amlekar led motion graphics, then programming and robotics, before serving as the club's president from 2017 to 2018. The titles sound very tidy for a teenager. The work was cheerfully less so. He taught C++ and Python workshops, helped run an inter-school computer-science symposium for more than 300 participants from over 16 schools, added three events to its existing eight, and started an intra-school version to bring more students into computing.
The same years supplied an eclectic collection of distinctions: fourth place among 10,880 competitors in a national cryptic hunt, an international rank of 43 in the National Cyber Olympiad, and a school gold medal for seven consecutive years of academic excellence. A renewable-energy paper was accepted by India's National Innovation Foundation. His interests did not so much occupy a lane as acquire a traffic system.
Then there were the small machines. Brailliant used buttons attached to a Raspberry Pi to accept Braille patterns and translate them into speech in real time. An encryption project applied an XOR cipher in C++. His Devpost portfolio lists four hackathon projects, all marked as winners, including an augmented-reality education tool, a project for landmine awareness and a determinedly unserious game called Goose Snake. The portfolio's linked GitHub handle is GodofDragons, which is difficult to improve upon as an early professional brand.
The range extended beyond software. Amlekar was a founding member and captain of his school's physics club, RedShift Physics, from 2014 to 2018. He also directed technology and design for a student organization and lists four languages on his public profile: English, Hindi, Marathi and French. None of those details predicts a career in machine learning by itself. Together they show someone repeatedly volunteering for the point where technical work had to meet an audience, whether that meant teaching a workshop, designing an event or explaining a project to people outside the codebase.
Two internship recommendations emphasize the same cluster of qualities: curiosity, creativity, positivity and an ability to learn quickly. One Kinaxis manager also singled out Amlekar's presentation of a machine-learning project to the executive team.
02 / Learning the interfaceWaterloo made the human part explicit
Amlekar entered the University of Waterloo in 2018 and graduated in 2023 with a Bachelor of Applied Science in Software Engineering. His specialization was human-computer interaction, a choice that gives the later machine-learning work a useful frame. HCI asks the awkward questions engineering can postpone but never escape. What does the person think is happening? Where does the interface make them hesitate? When the system is wrong, can they tell it so?
His capstone project was not an agent or a productivity tool. It was Kill the Lich, an asymmetric multiplayer dungeon crawler built by a five-person team. Three heroes advanced through rooms while a fourth player, the Lich, placed enemies and traps to slow or eliminate them. The design joined roguelike urgency with tower-defense strategy. One side acted; the other interpreted and responded. Even the fantasy villain needed a feedback loop.
The co-op calendar meanwhile offered a brisk tour of applied computing. At Kinaxis, he worked as a machine-learning developer. At Intact, he was a data scientist. Syngli put him in software development; Qualcomm in software engineering; Nuro in behaviour simulation. In 2021, he also spent six months as co-founder of an unnamed startup. Public descriptions do not reveal its product, but the episode adds an operator's interval to a record mostly written in code and experiments.
03 / The production lessonSix billion tokens of people getting things done
Web agents have an embarrassing problem: the web. A clean benchmark can make digital action look orderly, but real websites are crowded with menus, nested forms, shifting layouts and tasks that wander across many steps. General-purpose language models may be eloquent and still lose the plot halfway through an expense report. The ScribeAgent team, which joined Scribe engineers with researchers from Carnegie Mellon University, chose to specialize.
Amlekar and six co-authors fine-tuned open-source models on production-scale workflow data drawn from more than 250 domains. The corpus totaled six billion tokens. It emphasized the kind of context web agents need: HTML, actions and long sequences that preserve the relationship between what appeared on screen and what a person did next.
The results were concrete. ScribeAgent reported leading direct-generation performance on Mind2Web and improved task success on WebArena by 7.3 percentage points over the previous best text-only agents. The paper also tested decisions that are less glamorous and more useful to builders: model selection, fine-tuning recipes, context-window length and dataset size. The point was not merely to make an agent sound informed. It was to make the next action more likely to be correct.
The first version appeared as a preprint in November 2024. In 2025, it was presented at the ICLR workshop on Foundation Models in the Wild and published in Transactions on Machine Learning Research. The venues fit the project. ScribeAgent lives at the border between research and the untidy evidence of production use, where a promising result must eventually survive someone clicking the wrong tab.
04 / The return of correctionPatents mature slowly; ideas travel faster
While the ScribeAgent work was advancing, Amlekar's first machine-learning chapter reached a delayed conclusion. Two United States patents from the Kinaxis inventor group were granted in November 2024, and a third followed in March 2025. They describe variations on interactive machine learning: a system presents predictions, a user adjusts results or data, and those interventions can feed the model's next round of learning.
Patent grants are poor clocks for biography. They arrive years after the original work and can make an old idea appear suddenly new. Here the delay is clarifying. In 2019, Amlekar was working on a formal route for human correction to improve a model. In 2024, he was part of a team training web agents on records of human work. The scale and architecture changed. The respect for interaction did not.
A career can be coherent without being narrow. Sometimes the through-line is the problem a person keeps choosing.
It is tempting to tell stories about artificial intelligence as a procession of larger models. Amlekar's public record suggests a quieter account. Better systems also come from choosing the right evidence: a corrected prediction, a sequence of clicks, the context around an action, the friction visible in an interface. His school workshops, HCI specialization, game design, simulation work and agent research all put a person somewhere inside the mechanism.
That may be why the phrase “workflow data” matters more than its corporate plainness suggests. A workflow is a small theory of how someone gets from intention to result. Capture enough of them carefully, and software receives something textbooks rarely provide: examples of people solving actual problems inside actual constraints. For an engineer who began by letting users correct a model, it is an apt next chapter. The machine is still learning to listen. The room has simply grown much larger.