Hyderabad, India · Open to relevant opportunities

Computer vision systems that work in the real world.

I’m Mohd Shahid, a Computer Vision and Machine Learning Engineer who builds low-latency video analytics—from dataset preparation and model training through multi-camera deployment, event processing, and operations.

Download résumé ↓Email me View selected work ↓

REAL-TIME INFERENCEMULTI-CAMERA ANALYTICSEDGE & GPU PIPELINESDEPLOYMENT ENGINEERING

01 / Experience

Production-minded from model to event.

MAR 2025 — PRESENT

Computer Vision Engineer

Nabeh · Hyderabad, India

Building real-time airport video analytics systems that turn frames, tracks, and OCR results into reliable operational events.

Prepared domain-specific datasets and trained YOLO models for baggage and airport-resource detection, reaching 96.35% mAP.
Deployed video pipelines across 26 CCTV cameras and 7 DeepStream containers, with under-100 ms latency using Python, NVIDIA DeepStream, YOLO, and OpenCV.
Integrated NemotronOCR for real-time identification of ULDs, baggage carts, and transporters.
Built detection, multi-object tracking, ROI analytics, session management, and structured-event workflows.
Connected GPU inference and Python services with Kafka, Redis, Docker, and POSIX shared-memory IPC; added recovery, validation, deduplication, and multi-frame voting for noisy or occluded conditions.
Automated multi-service, camera-specific deployments using Docker Compose and Python-generated configuration.

02 / Selected work

Public work, with the details a résumé cannot hold.

Each project shows the problem, technical approach, and evidence of outcome—not just a technology list.

Live demo ↗

DEEP LEARNING · KAGGLE

Messy Mashup Music Genre Classification

Trained a 10-class PyTorch classifier on instrument-separated stems—drums, vocals, bass, and more—then designed training and evaluation around cross-song mashups, tempo variation, and added noise.

Result0.87498 Kaggle leaderboard score

Source ↗

FULL-STACK SYSTEM

Home Services Marketplace

A role-based marketplace for customers, service professionals, and administrators. I designed booking and assignment flows, REST APIs, relational models, asynchronous jobs, caching, and scheduled notifications.

OutcomeCoursework grade: S

Kaggle competition

MACHINE LEARNING

Bank Telemarketing Success Prediction

Built a binary classifier for term-deposit subscriptions using customer and campaign data. Evaluation focused on macro F1 to account for class imbalance, then informed model selection and tuning.

Result0.75004 Kaggle leaderboard score · Grade A

More public work: GitHub ↗ · Academic GitHub ↗

03 / Technical stack

Tools in service of reliable systems.

Computer vision & ML

PyTorch, YOLO, OpenCV, scikit-image, NVIDIA DeepStream, TensorRT, ONNX, NemotronOCR

Video analytics

Object detection and tracking, multi-camera analytics, real-time inference, image processing, ROI-based analytics

Systems & APIs

Apache Kafka, Redis, POSIX shared memory, IPC, FastAPI, REST APIs, SQL

Deployment

Docker, Docker Compose, Linux, Git, Python, Vue

04 / Background

Engineering judgment comes from building, breaking, and rebuilding.

I work where machine learning meets operational software: data and models are only useful when the surrounding pipeline can keep up with real environments, imperfect inputs, and people who depend on the output.

My curiosity also extends to communication, language, and human behavior. That perspective helps me approach a technical system as a whole—its constraints, failure modes, interfaces, and intended use—not only as a model.

B.Tech in Computer Science: Maulana Azad National Urdu University, Hyderabad · 2024
Diploma in Data Science: Indian Institute of Technology Madras · 2026
Diploma in Programming: Indian Institute of Technology Madras · 2025

05 / Contact

Have a computer vision, video analytics, or ML systems problem?

msaligs@gmail.com ↗

Project notes

The implementation behind the result.

These notes add the engineering context that would make a résumé too dense.

01 / AUDIO CLASSIFICATION

Messy Mashup Music Genre Classification

Challenge: Classify 10 music genres when instrument stems are recombined into cross-song mashups, with tempo variation and added noise—not clean, single-track audio.
Approach: Built a PyTorch training and evaluation pipeline around instrument-separated drums, vocals, bass, and other stems. Used audio features and experiment tracking to test how well the model generalized under the competition’s distorted conditions.
Evidence: Reached a 0.87498 Kaggle leaderboard score. The live demo is available through the project card above.

02 / FULL-STACK MARKETPLACE

Home Services Marketplace

Product model: Three role-specific experiences: customers create service requests, professionals manage assigned work and availability, and administrators manage users, professionals, categories, locations, services, and reporting.
Backend design: Implemented Flask blueprints and REST endpoints with role-based access control, SQLAlchemy relationships, booking-to-professional assignment, request-status workflows, reviews, categories, and location data.
Asynchronous work: Used Redis as the broker/cache with Celery workers and Celery Beat for background CSV exports plus scheduled daily and monthly service-status emails.
Delivery: Containerized the Vue frontend, Flask API, Redis, Celery worker, and scheduler with Docker Compose so the complete system can start as a coordinated local environment.

03 / TABULAR MACHINE LEARNING

Bank Telemarketing Success Prediction

Question: Predict whether a bank client will subscribe to a term deposit using client and campaign features.
Evaluation choice: Treated the target as imbalanced and optimized with macro F1 rather than relying on accuracy alone, so minority-class performance influenced model decisions.
Outcome: Achieved a 0.75004 Kaggle leaderboard score and an A grade. The project reinforced the connection between metric selection and the real decision a model is intended to support.