/ TMI is a social construct. Here's my life story.
01The original plan
I started in physics wanting to work in astrophysics.
02Plot twist
Partway through that degree, a Google Data Analytics course pulled me into data, and I began a second degree in data science at IIT Madras while finishing the first.
03Finding the signal
My first machine learning work stayed close to the sky: segmenting rocks in lunar imagery with a U-Net, then classifying 22 kinds of noise in LIGO's gravitational-wave detectors. Both were the same problem underneath, separating a real signal from everything that only looks like one.
04Following the signal
That problem took me to markets. At WorldQuant I research and backtest systematic equity signals across US, Chinese and Indian markets, where anything that only works in a backtest gets found out quickly. Alongside it I build with LLMs: agents with long-term memory, retrieval systems, and RL environments that teach models to repair code.
05Now: back to where it started
Research is bringing me back to physics. I work on physics-informed neural networks, particularly why they break down on stiff, multi-scale and transport-dominated problems, and what makes them train reliably.
∞The thing that never changed
What carries across all of it is evaluation. Whether it's a trading signal or a pile of messy, real-world data, the work is the same: clean it, test whether the pattern is really there against strong baselines and out-of-sample checks, and turn the answer into something a non-technical person can act on. Results that hold up, not ones that only look good.
Roles I'm drawn to: ML research, quant research, data science and analytics.
What I work with mostPython, SQL, PyTorch, LLMs and agents, statistics, time series, NLP
/ things I've built
My ten most recent projects, from multi-agent systems to a real business's data. Each opens into the full case study, and there's more on my GitHub.
/ education
Two bachelor's degrees, run side by side. Pick a bar for the details.
Beyond the degrees
- Google Data Analytics Professional CertificateTaken before I started the data science degree. It's what pulled me into the field in the first place.
- Cloud Computing and Data Mining, IIT Kharagpur (NPTEL)Two courses taken outside my degree; the credits transferred into my IIT Madras degree.
- AWS Cloud Practitioner coursework, HoneywellCloud fundamentals, completed with an A+, the highest grade.
- Advanced Diploma in Cyber LawsA field completely different from my two degrees, but a close neighbour of data science: the law of data privacy, data protection and cyber security, studied across three levels.
Outside the classroom
Member of the Women in Tech, Coders' High and Ramanujan Research societies at IIT Madras.
/ skills
The ones on my CV, and the ones that aren't. Both are honest.
/ experience
Research Consultant @ WorldQuant
- I design trading signals, rules that predict which stocks will do better than others, for the US, Chinese and Indian markets.
- I write them in Python, drawing on everything from company fundamentals and credit risk to news sentiment and machine-learning forecasts, with checks built in to catch look-ahead bias and data too short to trust.
- Rather than ten versions of the same idea, I combine signals from genuinely unrelated data, like credit risk with news sentiment, and strip out broad market risk so what's left is specific to each stock.
- My best signals have reached a backtested Sharpe ratio of 3.05. These are research results, not live trading or personal investment performance.
Data Analytics Extern @ Beats by Dre
- Built an NLP pipeline to score the sentiment of customer reviews for Beats and ten competing headphone brands, then tested which signals actually drive engagement. Point-biserial correlation showed that critical Beats reviews attract markedly more "helpful" votes (r = −0.36), twice the market-wide effect.
- Ran aspect-level analysis across price-performance, sound quality, durability, noise cancelling and customer service, comparing how reviewers rated Beats and its competitors on each.
- Synthesised the evidence into competitive positioning, a SWOT analysis and product and marketing recommendations. See the code.
Business Analytics Extern @ HP
- Worked with HP Tech Ventures, HP's venture capital arm, to find early-stage startups worth investing in.
- Built market maps of 50+ startups, collecting and cleaning the data with SQL, Python and OpenRefine, so the team could see who operated in each space and how they compared.
- Screened them against the fund's criteria: funding stage, investment size, US market presence, a working product and market size.
- Pitched the best fit as an investment, backed by market sizing, a competitor comparison and due diligence on the founding team and business model.
ML Research Intern and Trainee @ Spartificial
- Led a team building deep-learning classifiers for 22 classes of LIGO gravitational-wave glitches from Gravity Spy spectrograms, benchmarking VGG16, VGG19 with an attention module, and MobileNet under transfer learning.
- Ran systematic ablations over attention mechanisms, label smoothing, L2 regularisation, dropout and multi-branch feature extraction, with validation-based early stopping and checkpointing. MobileNet with label smoothing reached 89% test accuracy, 33 points above VGG16.
- As a trainee, built a U-Net semantic segmentation model for lunar terrain, with normalisation and augmentation designed for harsh lighting and low contrast, reaching 0.80 IoU and ranking first of 30+. Also built CNN classifiers for astronomical objects on SDSS survey data.
- Code: glitch classification and lunar terrain.
/ recognition
First, a quick estimation puzzle
Math puzzles are my favourite way to take a break, so here's one for you. I ranked #319 in WorldQuant's global Alpha Challenge. How many people do you think took part?
/ research
I'm currently working on a research project on physics-informed machine learning for sensor data, under the guidance of Dr. Bharavi Mishra. Alongside it, these are the questions in the field that have my attention.
Why physics-informed networks fail, and what fixes them
A physics-informed neural network (PINN) combines observations with governing physical laws, typically through a loss that includes data or boundary conditions alongside the residual of a governing PDE evaluated at collocation points. The premise is intuitive: let the underlying physics constrain what the model can learn when data is thin. Whether that actually translates to more reliable generalization in practice is a much more nuanced question.
The slider captures the central tension. Dial the physics weight down, and the model can fit the observations while paying less attention to the governing equation. Dial it up too high, and the PDE residual can dominate the optimization at the expense of the observations.
Three things make PINNs particularly interesting to me:
- Loss balancing: The data and physics terms can have very different scales and gradient behaviour, making fixed weights difficult to choose. This is why adaptive schemes that rebalance the terms using gradient statistics or neural tangent kernel (NTK) based approaches remain an active area of research.
- Spectral bias: Neural networks tend to learn low-frequency structure before higher-frequency detail. That can make sharp fronts and multi-scale behaviour difficult to capture. Fourier features and domain decomposition are among the approaches researchers use to address these difficulties.
- Collocation points: Where the model is asked to satisfy the physics matters. Uniform sampling can spend computation in regions that are already easy to approximate, while residual-based adaptive sampling puts more points where the PDE residual is large or the solution is difficult to capture. How to allocate those points efficiently is an important part of making PINNs work well when data is scarce.
What I'm following
- AI safetyEvaluation, interpretability, and control: how we tell whether a capable system is doing what we intended, and what to do when it isn't.
- Rules-aware MLPINNs and neural operators: models that incorporate physical constraints rather than learning entirely from statistical patterns in the data.
- AgentsArchitectures that give language models persistent memory, tool use, and ways to verify their own answers.
- Training environmentsRL setups designed to teach models concrete, verifiable skills, like repairing broken SQL, with automated scoring.
- Quant researchSignal decay, market impact, execution drag, and the gap between a backtest and live execution.
/ side quests
How I spend my evenings and weekends.
Hackathons and competitions
Meta × Scaler, a Google × Kaggle AI agents intensive, and a run of IIT Madras competitions in speech, NLP and generative AI. I've only done a handful so far, and every one sent me home with something worth keeping: most of the projects above started at one.
Kaggle
My machine learning practice ground since last year. Testing ideas against a real leaderboard, and reading how other people solved the same problem, has taught me more than most courses.
Open source and citizen science
Just getting started, and already hooked. I contribute to open-source projects and to citizen science projects on Zooniverse and NASA, where volunteers work on real research.
Off the clock
When the side quests run out, I read: fiction of almost any kind, from literary and contemporary to thrillers and horror. Non-fiction only makes the cut if it's about science. And when I start missing sunlight, I leave the house, travel a little and touch some grass.
/ say hi
If something here made you curious, or you just want to swap notes on a research idea or a messy dataset, my inbox is always open.
Or we can skip straight to the important stuff: book recommendations and chess challenges.