Shubhransh Jagota

Designed and built a natural language–controlled browser agent that autonomously performs multi-step web tasks — planning, navigating and acting across pages without a scripted path. Led the migration of the Actions Hub to a modular architecture, decoupling action registration from agent execution, and instrumented the LLM systems with Langfuse.

Author of the Java SDK for PhonePe's B2B payment gateway — the SDK family carried tens of crores in total payment volume. Owned the Epoch scheduling framework for online merchant services, and led the disaster recovery drills end to end with SREs, including the migration to the Mahape data centre. There's more under the PhonePe tag.

On the Fixed Income Division's Macro Valuation Adjustment team — a valuation adjustment being the regulatory capital a bank holds to satisfy its regulators, kept to just enough. Took 10+ of them off manual footing across desks and jurisdictions, tightening controls and saving ~25 engineer-hours a month. There's more in the write-up.

Built a cost-efficient heating solution using machine learning for Energy.AI clients, prototyping an approach built on genetically evolving algorithms. Projected a 40–50% cost reduction for clients.

Worked on Project PAL, which aimed to deliver personalized learning to its users. I built the reinforcement learning environment, wrote the Python simulations of user behaviour, and generated the augmented data used to train the algorithms.

I went to UNSW as a visiting student, and picked up work as a research assistant for the UNSW Business School alongside it. Under Professor Fethi Rabhi I built a system to assess the impact of events on assets, and contributed to research on dark pool and hedge fund activity at the ASX — analysing trading patterns and fund distribution across the Australian exchanges. Part of it is open source: RA_Bschool.

Run by Citadel and Correlation One, across a field of more than 30,000 engineers and data scientists. Represented MIT against Harvard as a team of two with Akash Vaish. On Citadel's Terminal you don't play the game — you write an algorithm that plays it against other people's algorithms. Ours — Go Get It, which the versions turned into go, go1, go2, go3 — adapted its offence to the opponent's attack. We made the final as one of three teams a side and lost 5–4. The full story.

One week I was getting bored. I saw a competition, decided to spend some time on it, and got lucky in the end. I was working through ML courses then and wanted to see if any of it had a practical use — logistic regression with hyperparameter tuning ended up performing best. The submission scored above 0.03, which put me in the top prize bracket, and it came with a cash prize.

You miss every shot you don't take. The full story.

Ranked 23rd internationally and 2nd in India, with a leaderboard score of 32,091. Alphas were written in WorldQuant's expression language and backtested in WebSim. The Alphathon led directly to a research consultant offer, which I held for two years. The full story.