Sports CV · Cricket Analyst · Self-documenting builder

About.

I'm a Sports CV engineer at KhiladiPro. By day I work on production computer vision for live cricket and badminton — annotation pipelines, ball and shuttle tracking, action recognition, biomechanics. By night I build public cricket analytics — a deployed ODI match predictor, IPL scouting models, fantasy XI optimization, venue intelligence. Two halves of one path: getting good enough to work on elite sport at the highest level.

01 / Where I Am

Sports CV Engineer at KhiladiPro.

— Day Work
Production computer vision for live cricket and badminton broadcasts. The pipeline goes from raw broadcast video → annotation engine → object detection and tracking → action recognition → biomechanics → coach-facing intelligence. I work across the stack, not on a single sliver of it. Real footage means real constraints — occlusion, motion blur, variable camera setups, noisy audio. Solving those is most of the job.
— Night Work
Cricket analytics, built in public. A deployed ODI Full Match Predictor with calibrated probabilities and temporal train/val/test splits. An IPL Player Scouting Model with K-Means role clustering. A Fantasy XI optimizer using Integer Linear Programming. A Venue Intelligence Hub. Some of these have ML bugs I left in and documented honestly — because knowing what a target-leakage signature looks like is more valuable than hiding it.
— Why Both
The CV work teaches the parts of the stack that are hard to learn from books — broadcast-grade pipelines, real-world data ingestion, deployment at production scale. The analytics work teaches the parts work doesn't yet touch — calibration, model selection rigor, optimization. Each half makes the other one better.
02 / How I Think About The Work

Four principles I'll defend.

— 01

Pipeline beats model.

80% of every project I've shipped was clean data, careful joins, version control, and parsing edge cases. The model is the last 20% — and the model gets all the credit while the pipeline does all the work. I'd rather have a clean dataset and a logistic regression than a dirty one and a transformer.

— 02

Match the model to the data.

Reaching for XGBoost on 2,000 rows is a tell. On the ODI predictor I compared Logistic Regression, Random Forest, and XGBoost on a temporal validation set — LR won at 68.8% accuracy because 13 features on 2,005 rows is too small for tree boosting. Choose models for their inductive bias, not their reputation.

— 03

Honest metrics or none.

A 1.0 accuracy is a signal to investigate, not celebrate. Calibration matters more than peak accuracy — a well-calibrated 60% beats a miscalibrated 90% for any downstream decision. If I can write the label as a function of the features in five lines of pseudocode, I don't have a prediction problem.

— 04

Document obsessively.

This portfolio exists because I write about every project I ship — including what's broken. Each project page has a Limitations section that lists real bugs by line number. The writing isn't decoration; it's the work that turns a notebook into something defensible.

03 / The Stack

What I use day-to-day.

— Computer Vision

Production CV at KhiladiPro

  • Python · PyTorch · OpenCV
  • YOLOv8 · TrackNet · MediaPipe
  • Pose estimation · Biomechanics
  • Custom annotation tooling
  • FFmpeg · video pipelines
  • Audio-visual event detection
— Analytics

Cricket models on the side

  • pandas · NumPy · scikit-learn
  • XGBoost · Gradient Boosting
  • K-Means · Quantile Regression
  • PuLP · Integer Linear Programming
  • Streamlit · Plotly · deployment
  • ICC-style Elo · isotonic calibration
04 / The Path

Where this is going.

The honest version: I want to work on elite sport at the highest level. IPL teams running production analytics. ICC tournament intelligence. Broadcasters whose win-probability widgets are calibrated. National team biomechanics labs. The places where the work I do at KhiladiPro and the work I do on my own time both become professionally useful at the same time.

I'm not there yet. The portfolio exists to compress the gap — to make the work I've done visible enough, and rigorous enough, that the question of whether I'm qualified becomes uninteresting. Each project is documented like it's going to be defended in an interview tomorrow — because at some point, it will be.

The plan is simple: keep shipping at KhiladiPro, keep building public projects that get harder, keep writing about both. The work I'll be doing in two years isn't the work I'm doing today, and I'm fine with that. The trajectory is the point.

05 / Selectivity

What I won't do.

The work I'll defend has the work I won't do as its foundation. These aren't preferences — they're the lines that make the rest of what I ship credible.

Open to conversation

If you're working on cricket or badminton CV at scale, I want to hear from you.

Broadcaster analytics, franchise data teams, ICC initiatives, sports-tech startups, academic biomechanics labs — the work that bridges raw broadcast video and coach-ready decisions. Drop a line.