Bhavya Sharma
CV
Illustrated portrait of Bhavya Sharma

Hi, I'm Bhavya.

I build models for messy, real-world data, and I care most about whether they still hold up outside the notebook. Right now that means systematic equity research at WorldQuant, building with machine learning and LLMs, and research on physics-informed neural networks.

See my work View my CV

About me

I started out in physics and moved into data science through a second degree at IIT Madras. What connects everything I work on is separating a real signal from noise, whether it's in gravitational-wave data, a business's books or the stock market.

Most of my work runs on Python and SQL, with PyTorch for deep learning, and statistical testing underneath all of it.

Experience

Research Consultant @ WorldQuant

Aug 2024 – present

  • Research and develop systematic equity signals ("alphas") for US, Chinese and Indian markets, across universes from the 200 to the 3,000 largest stocks.
  • Engineer signals in Python from fundamental, credit-risk, news-sentiment and machine-learning forecast datasets, with automated checks for look-ahead bias and insufficient data history.
  • Combine signals from independent data sources under statistical risk neutralisation, isolating stock-specific returns from common market risk.
  • Validate every signal for out-of-sample robustness, turnover, risk concentration and correlation with existing signals. My best signals reached a backtested Sharpe ratio of 3.05 and annualised backtested returns of up to 29%.
  • Ranked in the top 0.7% (319 of 47,839) in WorldQuant's global Alpha Challenge and the top 2% (806 of 47,000+) in IQC 2024.

Backtest results on historical data, not live trading or personal investment performance.

Projects

Best Project of 3,000+ at IIT Madras

Fixing a distributor’s cash flow

Late payments were quietly draining a profitable TV distributor’s cash, the same problem that sinks many small businesses. I predicted which dealers would pay late before they did, priced every credit policy with simulation (the popular early-payment discount loses money), and turned it into ten changes the owner can run with tools he already uses, plus a Power BI dashboard to track them.

  • Python
  • pandas
  • statsmodels
  • SciPy (HiGHS)
  • scikit-learn
  • Power BI
  • DAX

Read the full story ↗View the code ↗

1 of 3,000+Best Project, IIT Madras
0.83ROC-AUC for a credit scorecard using only invoice-date data
~60%of TV gross margin lost to slow payments, now measured

Hindi disfluency restoration

IIT Madras speech recognition competition

0.205 word error rate with the smallest Whisper model and no fine-tuning

Speech recognisers delete the hesitations and repeats of natural Hindi speech. This restores them by fusing Whisper output with a disfluency lexicon and language-model scoring.

  • OpenAI Whisper
  • scikit-learn
  • TF-IDF
  • Bigram LM

Read the full story ↗

Placing ancient inscriptions

IIT Madras NLP competition

0.447 macro-F1, 13.7% above a strong TF-IDF + LightGBM baseline

Classifying 120,000 short, noisy inscriptions in many scripts into 15 regions of origin.

  • XLM-RoBERTa
  • Hugging Face
  • PyTorch
  • LightGBM

Read the full story ↗

Mapping rocks on the Moon

Spartificial ML training

0.80 IoU, first of 30+ trainees

Pixel-level rock segmentation on lunar terrain, built for harsh lighting and low contrast.

  • TensorFlow
  • Keras
  • U-Net

Read the full story ↗

Want to see more projects? They're on my personal site ↗, with the full story behind each one.

Skills

  • Data and code

    • Python
    • SQL
    • pandas
    • NumPy
    • SciPy
    • statsmodels
    • matplotlib / seaborn
    • Git

  • Machine learning

    • scikit-learn
    • Gradient boosting
    • PyTorch
    • TensorFlow / Keras
    • CNNs
    • U-Net
    • GANs
    • Reinforcement learning

  • LLMs and agents

    • Google ADK
    • LangChain
    • Multi-agent systems
    • RAG
    • Vector search
    • ChromaDB
    • Hugging Face Transformers
    • PEFT / LoRA
    • Whisper

  • Statistics

    • Statistical inference
    • Regression
    • Bayesian statistics
    • A/B testing and power analysis
    • Monte Carlo simulation
    • Survival analysis

  • Time series and optimisation

    • Forecasting and backtesting
    • Stochastic processes
    • Linear and stochastic programming
    • Numerical methods

  • Shipping and BI

    • FastAPI
    • Flask
    • Docker
    • Google Cloud
    • Power BI
    • DAX
    • Power Query

Research

Research project

Physics-informed machine learning for electrochemical sensors

Under the guidance of Dr. Bharavi Mishra, The LNM Institute of Information Technology · Ongoing

This project explores how to make machine learning reliable for electrochemical sensors, the technology behind glucose meters and point-of-care diagnostics, drawing on physics-informed neural networks, transfer learning, federated learning and generative models.

  • Learning from little data

    Sensor experiments are slow and expensive, so datasets are small. Physics acts as a constraint, so models need far fewer measurements to generalise.

  • Holding up across setups

    Sensors vary between production batches, instruments and labs, and drift as they age. The aim is models that stay reliable when data comes from different setups, where standard ML quietly breaks down.

  • Learning without sharing data

    Medical and industrial data often can't leave the institution that collected it, so models need to learn across sites while the raw data stays private.

What's next?

Say hi

I'm looking for research roles in machine learning and quantitative finance, and data science and analytics roles. Always happy to talk through a research idea.

Email meView my CV

Curious about the person behind the work? My academic journey, side quests and everything I get up to in between live on my personal site ↗