Open to Work

hey there, i'm Harin.

MSc Robotics at the University of Birmingham specilizing in Machine learning, Computer vision and AI Engineering. Building End-to-end system from data processing, model development to deployment across real-world applications.

Email Me GitHub LinkedIn Download CV
Harin22 / daily activity
Less
More
fetching last activity…
YouTube

Day in a life at University of Birmingham

Aesthetic minivlog • UoB campus life

Watch 1k+ views
Medium
LLM GUIDE

Everything you need to know before getting started with LLMs

Prompts, fine-tuning & workflows — start here.

Read article 5 min

Currently_learning

Live

Click to see in detail

Stack_Overview

Languages

Python C++ SQL Dart

Machine learning & Computer vision

NumPy Pandas Scikit-learn PyTorch TensorFlow OpenCV Detectron 2 Roboflow LangChain W&B HuggingFace YOLO (v5/v8)

Infrastructure & Deployment:

Git/GitHub ONNX Runtime FastAPI Flask Docker AWS S3 PostgreSQL MLflow Streamlit Gradio Supabase Postman GitLab GitHub Actions CI/CD Vercel Render Cohere Chroma DB DVC (Data Version Control)

Robotics

ROS 2 Gazebo Unity Carla Webots Raspberry Pi

Flagship_Projects

Real-Time Bubble Detection Pipeline

Masters Dissertation project - 73%
  • Developed a real-time bubble detection system using deep learning and instance segmentation to detect, count, and measure air bubbles in fluid-based manufacturing processes.
  • Trained and evaluated YOLOv8n-seg, YOLOv8s-seg, and Mask R-CNN with OC-SORT, achieving 87.6% mAP@0.5 and 65 FPS with YOLOv8n-seg.
  • Optimized the model using ONNX Runtime INT8 quantization, reducing model size by 50% (7 MB to 3.5 MB) while nearly doubling inference speed with minimal accuracy loss.
  • Deployed edge deployment pipeline targeting Raspberry Pi 5, demonstrating the potential for low-cost, real-time quality inspection in advanced manufacturing.
Roboflow PyTorch YOLOv8 Segmentation Model Training Quantization ONNX Export Real-Time Inference Raspberry Pi 5 Edge Deployment End-to-End Deployment
GitHub repo

Vision-Based Robotic Sorting

Assignment project - 74%

Design and implement a vision-guided robotic pick-and-place system for automated fruit sorting using a Quanser QArm and Intel RealSense RGB-D camera.

  • Developed a YOLOv5-based object detection pipeline to classify strawberries, tomatoes, and bananas with 96.3% precision and 94.3% recall on custom datasets (12k+ images).
  • Implemented depth-based 3D localization, camera-to-robot coordinate transformation and inverse & forward kinematics to compute accurate grasp poses for real-world manipulation.
  • Built autonomous modes and integrating motion planning, gripper control and Simulink-based HIL testing to validate robot behavior before hardware deployment.

Future Work: Planned to integrate a conveyor belt and design a custom robotic arm and end-effector from scratch, enabling optimized industrial automation by connecting computer vision, machine learning and robotic manipulation for real world industrial purpose.

Open CV YOLOv5 Forward & Inverse Kinematics End effector Model Training & Evaluation Inference Optimization Depth Estimation Motion Planning Simulink MATLAB
6k+ Youtube views Technical Document.pdf

Central Bank Digital Currency Fraud Detection Engine:

Fullstack ML & Deep learning
  • Developed Random Forest and XGBoost classification pipelines with behavioural feature engineering, temporal train-test splitting and MLflow experiment tracking to detect suspicious transactions and evaluate model performance using precision, recall, F1-score, ROC-AUC and PR-AUC.
  • Implemented a GraphSAGE-based transaction fraud detection model using PyTorch Geometric, representing wallets as nodes and transactions as edges, with Neo4j for graph storage and querying and edge-level classification to identify potentially fraudulent transfers.
  • Built a unified FastAPI inference service integrating Random Forest, XGBoost and GraphSAGE models behind a single /predict endpoint, returning fraud predictions and probability scores through JSON requestst
XGBoost Fraud Detection End-to-End ML Pipeline Graph Neural Network Synthetic Data Generation Data Preprocessing Feature Engineering Evaluation Metrics Fast API Docker Containerization
GitHub repo

Vision Language Model for Construction Site Safety

Multimodel and Gen AI
  • Built a video-based safety monitoring prototype integrating a LoRA fine-tuned SmolVLM-500M with YOLO11 to detect workers and classify construction-site scenes as SAFE, WARNING, or DANGER
  • Fine-tuned the VLM with PEFT/LoRA (1.1M trainable parameters, 0.22% of the model) on a custom 46-image dataset covering working at height, PPE compliance, and slips, trips, and falls, using fp16 and gradient checkpointing on a 4 GB GPU
  • Developed a video inference pipeline that runs YOLO on every frame and the VLM every 2 seconds, and exports an annotated video with detection boxes and verdict banners
Vision-Language Model (VLM) SmolVLM-500M Local model LoRA / PEFT Multimodal AI Hugging Face Transformers Video Inference Pipeline Parameter-Efficient Fine-Tuning On-Device / Edge AI
GitHub repo

Routeguard-ai

MLOps & RAG
  • Built a Streamlit app that analyses shipping routes between 25 ports, sampling 15 monitoring zones per route (SeaRoute) and pulling 24-hour forecasts from Open-Meteo
  • Developed a scikit-learn Logistic Regression pipeline using historical weather and GDELT-derived geopolitical features to estimate hazard probability.
  • Built a RAG pipeline (NewsData.io → Cohere embeddings → ChromaDB, Langchain) powering an LLM chat assistant and a geopolitical risk score, with Mercury/Gemini providers.
  • Risk-triggered rerouting: when a route’s risk exceeds the threshold, the app generates alternative routes by distance. Set up GitHub Actions CI/CD (lint/tests) and continuous deployment on Streamlit Community Cloud, with logging of route-analysis runs.
Feature Engineering Logistic Regression API Data Ingestion Risk modelling Geospatial analysis Chroma vector DB Cohere Embeddings LangChain RAG Pipeline LLM Chat Assistant GitHub Actions CI/CD Streamlit
GitHub repo Site is live

FixFlow — Code Debugging tool for vibe coders

SaaS
  • Built a full-stack AI debugging platform using React and Flask, helping developers identify coding errors, understand root causes, and generate corrected code using InceptionLabs Mercury 2.5.
  • Developed REST APIs for authentication, code analysis, and debugging history using Python and Flask, with JWT authentication and secure password hashing.
  • Integrated PostgreSQL via Supabase and implemented a three-analysis free tier with Stripe Checkout for Pro subscription payments.
  • Deployed the React frontend on Vercel and Flask backend on Render, delivering a cloud-hosted AI SaaS application.
React Flask PostgreSQL JWT Stripe payment Vercel Render Fast API Docker Containerization
GitHub Repo Fixflow is Live, Click to Try!

Mini_projects

Brain MRI Segmentation

Implemented and evaluated classical computer vision–based segmentation methods (Sobel, Canny, thresholding, and K-means) on 3D brain MRI volumes to isolate anatomical regions. Performed slice-wise and volumetric segmentation, and quantitatively compared methods using Dice Similarity, IoU, Pixel Accuracy, and Adjusted Rand Index (ARI) against ground-truth labels to identify the most effective approach.

Python OpenCV Scikit-learn Image Registration Medical Image Analysis Evaluation Metrics
Technical Report

Hackathons

Paper Pilot

Google_Hackathon_2025

Paper Pilot is a Chrome extension that helps students and researchers work with academic PDFs directly in the browser. It extracts text from research papers and uses built-in browser AI APIs to generate concise summaries and simplify complex academic content, improving readability and productivity without leaving the PDF view.

The extension integrates PDF.js for reliable text extraction, Chrome Extensions APIs for tab and side-panel management, and Chrome’s experimental AI APIs for in-browser summarization and rewriting. Designed with a clean, professional interface, Paper Pilot works across different PDF viewers and embedded documents while keeping all processing within the browser.

GitHub repo YouTube Demo

Career_Log

Computer Vision Research Support Assistant

University of Birmingham

Aug 2025 – Dec 2025
  • Collaborated with PhD researchers on Taylor bubble flow analysis, developing a custom Mask R-CNN–based segmentation pipeline achieving ~88.6% mAP for precise bubble detection in fluid dynamics experiments.
  • Built an end-to-end computer vision system combining instance segmentation, OC-SORT multi-object tracking, contour analysis, and edge-based slug length measurement, enabling automated extraction of bubble metrics.
  • Deployed a Gradio-based analysis web app for CSV export of bubble radius, width, and slug length metrics for research analysis.
Computer Vision Pipeline Deep Learning Segmentation Models Mask R-CNN OC-SORT Contour Extraction Model Inference Gradio Interface
GitHub repo

Data Analyst Intern

T-FAB Technologies LLP

Jun 2022 – Nov 2022
  • Queried SQL to extract, aggregate, and validate purchase and operational data from company databases, and performed exploratory data analysis and data wrangling to identify trends in sales and raw material
  • Built interactive dashboards using Excel and Tableau to visualize equipment utilization and cost-related metrics for reporting and analysis.
Excel SQL Pandas Matplotlib Seaborn Beautiful Soup Tableau

Engineering Intern

Hyundai India

Jan 2022 – May 2022

• Assisted in the development and testing of automotive cruise control systems, including sensor calibration and embedded ECU integration for real-world vehicle testing.

Teamwork Hardware–Software Integration Sensor Calibration System Validation Communication

Academic_Stack

M.Sc. Robotics

University of Birmingham

Sept 2024 - May 2026
Courseworks
Machine Learning Computer Vision Natural Language processing Robotic Path Planning Forward & Inverse Kinematics

B.E. Mechanical Engineering

Karpagam Academy of Higher Education

Aug 2018 - April 2022
Foundation_Stack
Linear Algebra Calculus & Stat Robotics Engineering Graphics Programming (C / Python) Mechatronics

Lets Connect!

Medium

Technical blogs on ai and robotics.

Behance

My UI/UX Designs.

YouTube

Love to documentation my journey.

Discord

Community hub.

About me!

Harin

I’m an engineer who enjoys building things end to end, from understanding the problem to deploying systems in the real world.

I’m a first-principles thinker who likes to wear multiple hats, especially at the intersection of AI and Robotics.

Curiosity and creativity is what drives me and my interdisciplinary background gives me an edge in different perspective and an advantage in connecting ideas across domains!