// running inference… 1 object detected
Bhavin
Dhedhi
I build machine learning systems that see — and the platforms that keep them running in production. Lanes, license plates, birds, fraudulent transactions: if it can be detected, I've probably tried to detect it.
// 01 · about
From antennas to attention
I started out in electronics — designing and fabricating wide-band microstrip antennas in Mumbai as an undergrad at KJ Somaiya College of Engineering. Somewhere between signal processing and a first convolutional network, I got hooked on teaching machines to perceive the world.
That thread ran through everything since: published research on license plate recognition and bird detection, self-driving car projects, a masters in Electrical Engineering at Columbia University focused on machine learning and deep learning, and production fraud-detection and vision systems at scale.
These days I work on data platforms — the infrastructure layer that decides whether ML in production is a superpower or a liability. Models are the easy part; the pipelines, feature stores, and feedback loops around them are where the real engineering lives.
// 02 · work
Trajectory
2022 — PRESENT
Senior Machine Learning Engineer, Data Platforms
- Building the data and ML platform layer powering models across the business.
- Previously ML Engineer II: shipped production fraud-detection systems and computer-vision models.
2021 — 2022
MS, Electrical Engineering
Columbia University · New York
- Focus on machine learning and deep learning; course assistant.
2018 — 2021
Software Engineer, Machine Learning
Think360.ai · Mumbai
- Credit-scoring models at 85% accuracy serving 100,000+ customers with instant loan decisions.
- Anomaly detection on petroleum-industry time series (SVC, random forests, PCA, Hotelling T²).
- Deep-learning POC that detected no-balls in cricket — YOLOv3 tracking the bowler's foot against the crease.
- R-Shiny decision tools for major pharma companies, cutting weeks of contract analysis down to minutes.
2014 — 2018
B.Tech, Electronics & Communication
KJ Somaiya College of Engineering · Mumbai
- Designed and fabricated two wide-band microstrip patch antennas (2–10 GHz).
// 03 · research
Published work
Automatic License Plate Recognition using Deep Learning
A 4-stage pipeline — three YOLO stages plus a CNN — that reads license plates in the wild, skew and bad lighting included. 82% end-to-end accuracy on unseen test images, with individual stages above 90%.
IEEE · 2018Detection of Birds in the Wild using Deep Learning Methods
Comparative study of YOLOv2, YOLOv3, and Mask R-CNN for detecting birds against complex natural backgrounds — relevant to aviation safety and ecological monitoring. YOLOv3 took the crown at F1 = 0.87.
// 04 · projects
Detections
graphicalForest
Extracts names, addresses, items, and totals from any invoice layout by combining adjacency matrices with random forests. Structure as a feature.
Advanced Lane Detection
Camera calibration, bird's-eye perspective transforms, and curvature math to track curved lanes in real-world units from dashcam video.
Let It Drive
Nvidia's end-to-end self-driving CNN, implemented to predict steering angles from raw camera frames — with on-the-fly augmentation via generators.
Traffic Sign Classification
LeNet, trained to read German traffic signs, with a heavy dose of data augmentation to survive the real world.
Lane Line Detection
The classic pipeline: edge detection, Hough transforms, and regression to collapse noisy segments into two clean lane lines.
Extended Kalman Filter
Sensor fusion in C++ — predicting position and velocity of moving objects from noisy LIDAR and RADAR streams.
// 05 · writing
Recent notes
Say hello
Interesting ML problems, data platform war stories, or just to compare notes on making models behave in production — my inbox is open.
bhavindhedhi96@gmail.com