The Minds Behind the Models: Kanishk Kumar Sachan

Kanishk Kumar Sachan: The AI Engineer Who Wants to See It Live

Kanishk Kumar Sachan didn’t come into quantitative finance through a traditional finance pipeline. He came through computer vision and robotics.

Currently completing his MSc in Artificial Intelligence at the University of Manchester, Kanishk’s background is in building production AI systems—from a 3D reconstruction pipeline for mining equipment at FLSmidth to an on-device health prediction app built in Flutter. He is also a Quantitative Research Consultant at WorldQuant BRAIN, where he builds automated alpha generation pipelines.

His entry into quant finance wasn't a calculated career move; it was a search for a better testing ground.

"I came into the Manchester MSc through computer vision and robotics, not finance. Somewhere in there I started building a reinforcement learning trading agent, partly just to get RL onto something with real feedback instead of a toy Gym environment. Watching a PPO agent make actual, initially terrible, trading calls is what hooked me."

That side project eventually became his dissertation, which investigates how agent architecture—memory design, role specialization, and risk control—influences trading behaviour in LLM-based systems. But the real turning point was a competition.

"IMC Prosperity 4 sealed it. Finishing well above benchmark against people who’d clearly done this a lot longer than me hits differently than a grade on an assignment."

HYDRA: Live Money, Live Markets

When asked about a project he’s excited about, Kanishk pointed to HYDRA—a live intraday trading bot running on Alpaca and hosted on DigitalOcean.

The strategy itself is a sophisticated blend: an opening-range breakout paired with VWAP-reversion, filtered through a VIX-based regime check, and sized using a Kelly-derived allocation model rather than flat position sizing.

But the strategy isn't what excites him most. The infrastructure is.

"What I actually care about is that it’s live money in live markets. A backtest survives everything. A volatile Tuesday morning does not."

This is a philosophy we've seen echoed across this series. Kushpreet mentioned iterating based on live PnL rather than backtests alone. Alok and Mathurin both built sophisticated systems to calibrate their local backtests against the reality of the server. Kanishk takes it one step further by deploying to the real market.

In a field where it's easy to fool yourself with a pristine equity curve, the willingness to let a model trade real money—and fail in real time if it's wrong—is a powerful forcing function.

Avoiding the Crowd

For the AlphaNova competition, Kanishk applied a principle he uses on the WorldQuant BRAIN platform: stay away from crowded trades.

"I stay away from the momentum and mean-reversion signals everyone’s already mining and dig into cross-asset and fundamental data that gets less traffic. By the time a signal’s crowded, the alpha’s usually already gone. I’d rather sit with noisier, underused data and put my effort into getting the risk-adjusted sizing right, instead of squeezing more out of something everyone else has already picked over."

This is exactly the kind of thinking our competition architecture is designed to reward. Our greedy quality selection process filters out signals that correlate too highly with what's already on the board.

If everyone is mining the same price-volume data, they produce signals that look like copies of each other. By deliberately shifting to cross-asset divergence and analyst estimate revisions, Kanishk was able to find novelty where others weren't looking.

It's a structural insight that we've explored in our research on signal diversity and the geometry of forecast libraries—the space of uncorrelated signals is vast, but you have to actively look outside the popular datasets to find it.

The Architecture of AI: Less Prompting, More Context

Kanishk uses AI constantly—he runs an automated alpha generation pipeline using Claude Code for his work at WorldQuant—but his perspective on how to use it is remarkably mature.

"Claude Code runs my alpha research pipeline end to end. I use LLMs for literature synthesis on the dissertation too, which is a bit funny given the dissertation is literally about agent architectures for trading. For hackathons I lean on newer agent-first tooling, Google Antigravity mostly, to move fast on a 24-hour clock."

But like several other top performers we've spoken to, Kanishk has realized that the AI tools are not the bottleneck. The architecture around the AI is.

"The alpha ideas were never really the hard part — structuring a CLAUDE.md that stops the agent from losing context, or re-running an expression it already tried and failed three sessions ago, is the actual work now. Less prompt engineering, more architecture."

He's touching on one of the most important practical skills in modern AI-assisted research: context management. If you are running a pipeline that tests hundreds of hypotheses, the AI needs to remember what it has already done so it doesn't go in circles.

This mirrors Vibhav's use of sub-agents to keep different stages of work isolated, and Alok's use of AI to keep a versioned record of what had already been disproven.

"Code generation is honestly the least interesting part of it. What actually helps is having something hold more context than I can across a long session and catch what I’d otherwise miss."

What We Can All Learn

Kanishk’s profile offers three highly actionable lessons for anyone building trading systems today:

A backtest is not reality. Deploying a strategy live—even with small size—exposes you to the regime shifts, slippage, and volatility that a historical simulation smooths over.
Avoid the crowded trades. If everyone is mining momentum and mean-reversion, your alpha will decay before you can extract it. Look at cross-asset and fundamental data that gets less attention.
Don't just prompt AI—architect around it. The most valuable AI skill today isn't writing the perfect prompt. It's designing the memory, context, and role structure that allows an agent to work effectively over long, complex sessions.

That combination of engineering rigour, strategic originality, and modern AI orchestration is exactly what the AlphaNova community is built on.

Stay tuned for more profiles from the AlphaNova community. If you’d like to be featured, reach out—we’d love to share your story.