Profile Picture

Umar Shafiq Ahmad

Data Scientist | AI,ML | Computer Vision


💻 Tech Stack: Python · OpenCV · TensorFlow · PyTorch · Scikit-learn · Deep Learning · Computer Vision · AI · ML · Transformers · LangChain · LangGraph · Generative AI · CI/CD Pipelines · API Development · EDA · ETL
🧰 Tools: MS SQL Server · Power BI · Git · GitHub · GitHub Actions · Azure · Docker · Advanced Excel
🌟 Soft Skills: Analytical Thinking · Strategic Planning · Positive Attitude · Time Management · Problem Solving · Attention to Detail · Service Quality Improvement · Team Collaboration

Experience

Data Analyst - Hitachi Payment Services Pvt. Ltd.

July 2024 - June 2025 💼 Full time

🔹 Automated Cash Logistics Optimization: Developed a Python-based cost analysis scheduler that crunched through two months of Sunday and holiday indent data, smartly categorizing manual vs. system-generated cases. This led to a 15% cost reduction and ~10% drop in avoidable CIT indents, thanks to automated root-cause detection and interactive HTML-based reporting dashboards.
🔹 Real-Time ATM Fault Diagnostics: Built Python diagnostic tools to detect faulty cassette configurations across live ATMs, triggering instant alerts for the ops team. This innovation drove a 15%+ cut in ATM cash-out incidents, faster fault resolution, fewer penalty cases, and a clear boost in revenue retention.
🔹 Forecasting for Smarter Operations: Engineered ATM evacuation forecast scripts that automated 10-day and 30-day cash planning, factoring in weekday/weekend trends and rolling dispense data. This eliminated ~40% of manual Excel work and significantly accelerated operations planning cycles.
🔹 Data Reliability & Pipeline Monitoring: Monitored ETL workflows to maintain >80% on-time data refreshes for the SBI account, reducing operational risk by ~25%. Also implemented automated root-cause analysis tools to diagnose and prevent forecast failures before they impacted business decisions.

ML Engineer - Ramdisk.io

Jan 2023 - June 2024 💼 Full time

🔹 Real-Time Object Detection: Deployed and fine-tuned YOLO models for desktop-based, real-time object detection using Python and OpenCV - achieving 75%+ accuracy with sub-200ms latency on live video feeds. Enabled smooth, real-time analytics for camera-based applications.
🔹 Edge AI for Safety Monitoring: Developed and integrated a TensorFlow Lite-based helmet detection system on Raspberry Pi, optimized for low-power environments. Reduced safety violation detection time by ~40% in real-world conditions, enabling rapid, on-device decision-making.
🔹 On-Device Image Classification: Built custom TensorFlow models for multi-class image classification in Android applications, delivering >85% accuracy through efficient on-device inference. Streamlined deployment pipelines for scalable edge-AI applications.

Projects

🎬 Local RAG Chatbot

A fully local AI chatbot that understands and talks about movies - powered by a Retrieval-Augmented Generation (RAG) pipeline built from scratch.
Runs completely offline using TinyLlama (1.1B) — optimized to fit on an NVIDIA 1650 (4 GB VRAM) setup.
Uses Chroma Vector Store + Sentence Transformers to perform lightning-fast semantic retrieval on local movie data.
Built with LangChain to handle contextual reasoning, prompt flow, and grounded responses.
Can answer natural queries like “List Nolan's top-rated thrillers” or “What's the IMDB rating for Inception?” - all from your local dataset.
Designed for privacy-first and low-resource AI applications, showing how advanced LLM workflows can thrive on modest hardware.
Scalable beyond movies — adaptable for books, company wikis, or offline knowledge assistants.
🧠 A glimpse of the future where powerful AI stays local, private, and personalized.

Github

🧭 AI-Powered Defect Detection & Quote Generation

A dual-purpose FastAPI application showcasing computer vision and NLP capabilities.
Features a YOLO-based defect detection system that identifies and localizes 6 types of surface defects (crazing, inclusion, pitted surface, patches, rolled-in scale, scratches) with bounding box visualization.
Includes a custom-built language model using GPT-2 architecture, trained from scratch on TinyShakespeare and fine-tuned with LoRA on motivational quotes dataset for keyword-based quote generation.
Deployed using Docker containers on Render with full REST API documentation.
Demonstrates end-to-end ML engineering from model training to production deployment.
Tech Stack: FastAPI, PyTorch, YOLO, Transformers, LoRA, Docker, Render

Github

🛡️ AI SmartGuard

An end-to-end AI-on-the-edge simulation system that automates the entire pipeline - from training and deployment to testing - just like running on a real embedded device.
Trains lightweight AI models for real-time tasks like helmet detection, gesture recognition, and object classification.
Converts and optimizes models to TensorFlow Lite with full quantization support for edge efficiency.
Simulates embedded deployment locally - mirrors real-world environments like Raspberry Pi or Jetson Nano.
Enables MQTT-based inference control, letting you interact with models as if they're deployed on IoT devices.
Includes automated testing and benchmarking for accuracy, latency, and throughput - all integrated into a CI/CD pipeline.
Uses GitHub Actions to retrain, convert, deploy, and test on every code push - achieving complete MLOps-style automation.
Designed for edge-AI prototyping, industrial safety systems, and smart surveillance workflows where reliability meets low-latency AI.
⚙️ A living demo of how AI models can be built, optimized, deployed, and validated - all without ever leaving your laptop.

Github

Open to opportunities, Let's build something extraordinary together 🚀

📩 umarsunny2011@gmail.com
🔗 https://www.linkedin.com/in/umar-sunny/
🐙 https://github.com/umarsahmad