ML Engineer & Data Scientist
building Computer VisionDeep LearningGeospatial AIReal-Time Systems

Open to opportunities

OmkarKadam

Scroll to explore19.0760° N, 72.8777° EIIT Bombay · SIH ’24 Winner

(01) Selected work

Systems that
see, learn & scale.

[ 03 case files · 03 more · 04 lab ]

01 / 03Geospatial AI · ML SystemsIIT Bombay

AI Mining Intelligence Platform

Case file

AI-driven geospatial intelligence platform built through research at IIT Bombay. Ensemble ML pipelines, magnetic inversion and interactive 3D geological visualization help locate mineral-rich zones in large drill-hole survey datasets.

F1-score, up from ~30%: 70%+
ML models in ensemble: 5
Voronoi compute cut: ~30%
3D point-cloud viewer: WebGL

XGBoost
LightGBM
CatBoost
Random Forest
SMOTE
Magnetic Inversion
FastAPI
Docker
WebGL
Python
Octree

DhatuVerse — 3D geological visualization dashboard — FIG. 01Hover to inspect ⌕

The challenge

Predicting mineral prospectivity from sparse, noisy, heavily imbalanced borehole data — where a naive model scores around 30% F1 and classical GIS tooling breaks at scale.

The approach

Random Forest, Gradient Boosting, XGBoost, LightGBM and CatBoost ensembles with SMOTE-Tomek resampling and Stratified K-Fold / LOOCV validation, plus Octree-based spatial indexing to cut Voronoi computation by ~30%.

02 / 03Computer Vision · Deep Learning

Planetary Crater Detection

Case file

Automated crater detection for planetary surface analysis. YOLOv12 trained on 5,000+ satellite images with domain-specific augmentation, served as a Dockerized REST API for scalable orbital imagery workflows.

mAP score: 87%
Inference latency: ~200ms
Training images: 5,000+
Architecture: YOLOv12

YOLOv12
OpenCV
Python
Transfer Learning
Docker
REST APIs
Data Augmentation

AstroPit — YOLOv12 crater detection on planetary imagery — FIG. 02Hover to inspect ⌕

The challenge

Low-contrast terrain, ambiguous shadow edges and scarce training data — the boundary cases where standard detectors fall apart.

The approach

Transfer learning from aerial imagery, a custom augmentation pipeline for low-contrast features, and a containerized inference API tuned for low latency.

03 / 03Real-Time CV · HCI

Gesture-Controlled Virtual Drone

Case file

Webcam-based hand tracking that flies a virtual drone. MediaPipe landmarks feed a lightweight gesture classifier, rendered live in a Three.js simulation — no hardware controller required.

Gesture accuracy: 95%
Sustained real-time: 30 FPS
Full-pipeline latency: <100ms
Gesture controls: 8

OpenCV
MediaPipe
Three.js
Python
Gesture Recognition
Real-Time Systems

NeoPilot — MediaPipe hand tracking driving a 3D drone simulation — FIG. 03Hover to inspect ⌕

The challenge

Sub-100ms end-to-end control at a steady 30 FPS: capture, landmark detection, classification and 3D state update all inside one frame budget.

The approach

A trimmed MediaPipe landmark graph, a temporal-window classifier to kill jitter, and delta-only state sync between the Python pipeline and the browser.

(02) More work

Also shipped.

Hover or tab through the list

Best Paper · IJSRSET 2026

Multimodal deep-learning framework for real-time detection and early characterization of Earth-directed Halo CMEs from Aditya-L1 data. Fuses VELC coronagraph imaging with SWIS-ASPEX particle streams through temporal alignment, with reproducible space-weather benchmarks. Published in IJSRSET Vol. 13, Issue 8.

SIH 2024 Winner

AI-driven water resource optimization platform. Led a 5-member team through architecture, AI integration and delivery in the 36-hour national final of Smart India Hackathon 2024 — and won.

Machine Learning
Predictive Modeling
Team Leadership

Detect, track and maintain object identity across live video frames with stable bounding-box updates and optimized frame extraction.

Python
OpenCV
Video Processing

(03) The lab

Experiments & earlier builds.

01 / 04Face Emotion RecognitionScroll · clicktap a card

(04) About

I build intelligent systems that make sense of the world — networks that detect craters on planetary surfaces, read emotions, and understand hand gestures.

Mumbai, IndiaSIH 2024 Winner

Data scientist and ML engineer with expertise in predictive modeling, statistical inference and geospatial machine learning — from two years of research at IIT Bombay to an award-winning space-weather paper and a Smart India Hackathon win. I like problems where the data is messy and the stakes are real.

My approach is grounded in rigorous analysis, systematic experimentation and continuous learning. Good models are born from clean data, honest evaluation, and a willingness to question every assumption.

5Internships & research roles

70%+Mineral F1 · up from ~30%

87%mAP · AstroPit

500+Students mentored · TechNova

Certifications

IBM Watson StudioML model development & deployment
Microsoft CopilotApplied Generative AI
Deloitte Data AnalyticsVirtual Experience · Forage

GitHub LinkedIn

(05) Capabilities

My toolkit, in play.

Grab one · throw it · bigger = used in more places

XGBoost^×2Used at IIT Bombay, Yuva Intern · NSDC
Random Forest^×2Used at IIT Bombay, Yuva Intern · NSDC
Gradient Boosting · LightGBM · CatBoost^×1Used at IIT Bombay
SMOTE · class imbalance^×1Used at IIT Bombay
Cross-validation^×2Used at IIT Bombay, Yuva Intern · NSDC
Feature engineering^×2Used at IIT Bombay, Yuva Intern · NSDC
ML pipelines^×2Used at IIT Bombay, Edunet Foundation
OpenCV^×3Used at AstroPit, NeoPilot, Object Tracking
YOLO^×1Used at AstroPit
CNN · ResNet^×3Used at Edunet Foundation, Vault-Tec Security, Halosight
Multimodal sensor fusion^×1Used at Halosight
MediaPipe^×1Used at NeoPilot
Transfer learning^×1Used at AstroPit
Magnetic inversion^×1Used at IIT Bombay
Interpolation^×1Used at IIT Bombay
Orebody & lode modeling^×1Used at IIT Bombay
Point-cloud processing^×1Used at IIT Bombay
Python^×4Used at IIT Bombay, AstroPit, NeoPilot, Object Tracking
FastAPI · REST APIs^×2Used at IIT Bombay, AstroPit
Docker^×2Used at IIT Bombay, AstroPit
Power BI^×1Used at Yuva Intern · NSDC
WebGL^×1Used at IIT Bombay
Three.js^×1Used at NeoPilot
SQLWorking knowledge
CWorking knowledge
JavaScriptWorking knowledge
GitWorking knowledge
JupyterWorking knowledge
PandasWorking knowledge
NumPyWorking knowledge
SciPyWorking knowledge
Scikit-learnWorking knowledge
TensorFlowWorking knowledge
KerasWorking knowledge
PyTorchWorking knowledge
MatplotlibWorking knowledge
PlotlyWorking knowledge
SeabornWorking knowledge
TableauWorking knowledge
ExcelWorking knowledge
Statistical analysisWorking knowledge
Hypothesis testingWorking knowledge
A/B testingWorking knowledge
ETL pipelinesWorking knowledge
Data versioningWorking knowledge
Object detectionWorking knowledge
Image segmentationWorking knowledge

Source: resume and each project’s listed stack. Filled capsules = used in 2+ places.

(06) Career

Four roles, one trajectory.

Scroll to play the timeline ↓

Apr – May ’25

Jun – Jul ’25

Aug ’25 – Jun ’26

Jun ’26 – Now

01 / 04

AI & Data Analytics Intern

Edunet Foundation · AICTE

Apr – May 2025

Built end-to-end ML pipelines on 5GB+ datasets for training and deployment
Developed a ResNet50-based computer vision model with 89% accuracy

89%ResNet50 vision-model accuracy

02 / 04

Junior Data Scientist

Yuva Intern · NSDC

Jun – Jul 2025

Built Random Forest and XGBoost classifiers reaching 88% F1-score on 200K+ records
Engineered 50+ features and tuned hyperparameters with cross-validation, lifting performance by 18%
Developed Power BI dashboards that turned predictions into business decisions

88%F1-score on 200K+ records

03 / 04

Research Intern

Machine Intelligence Group · IIT Bombay

Aug 2025 – Jun 2026

Built data-preparation pipelines (outlier detection, imputation, coordinate normalization) and ran EDA on large-scale drill-hole records, surfacing distribution shifts
Developed interactive visualizations that made spatial and geological datasets explorable
Initiated the DhatuVerse ML workflow with evaluation benchmarks, baseline metrics and reproducible tracking
Standardized data versioning and quality-control checks across every geological dataset the group uses
Delivered orebody modeling, ore prediction, lode detection and alpha-beta analysis with regression and clustering models

10 moof geoscience ML research at IIT Bombay

04 / 04

Research Assistant

IIT Bombay · DhatuVerse

Jun 2026 – Present

Architected end-to-end ML pipelines for mineral prospectivity prediction, ranking Random Forest, Gradient Boosting, XGBoost, LightGBM and CatBoost ensembles
Countered class imbalance with SMOTE, SMOTE-Tomek and random undersampling, validated under Stratified K-Fold and LOOCV
Lifted F1 from ~30% to 70%+ through feature engineering and tuning, with significance tests confirming gains on held-out survey blocks
Built a magnetic inversion model recovering 3D subsurface susceptibility via regularized least squares, served through Dockerized FastAPI endpoints
Implemented isotropic, anisotropic, cubic and randomized interpolation for grade estimation across sparse drill-hole grids
Engineered a WebGL point-cloud system for large geological models with cross-sectional slicing

70%+mineral F1-score, up from ~30%