# Mahboob Alam

Source: https://alamm.me/

This is Mahboob Alam's self-authored public portfolio. Project descriptions and results are reported by the author; linked repositories and write-ups provide supporting details. Research, education, and personal projects are listed separately and do not imply employment at the organisations behind the tools used.

## Summary

I’m a machine learning engineer with an M.Tech in Data Science, specializing in computer vision and large language model systems. My work experience includes YOLO-based object detection, custom segmentation architectures, PEFT-based LLM fine-tuning, and RAG with Qdrant for document-grounded question answering. I work across the full model lifecycle, from data preparation to inference, using Python, PyTorch, and scikit-learn, and deploy models with FastAPI and Docker, testing through GitHub Actions. I’m motivated by understanding why models fail, tracing errors to their root causes to drive targeted improvements. Whether building vision models or LLM applications, I focus on measurable results and real computational constraints.

## Education

- M.Tech in Data Science | Defence Institute of Advanced Technology (DIAT-DU) | Pune, India | 2024 – 2026 | CGPA 8.13 / 10 · Coursework: Deep Learning, Machine Learning, Data Analysis
- B.Tech in Computer Science and Engineering | Rao Birender Singh State Institute of Engineering and Technology (IGU) | Rewari, India | 2019 – 2023

## Research Intern - Defence Institute of Advanced Technology (DIAT-DU)

MoES-Sponsored Research (Thesis-Integrated) | Pune, India | May 2025 – Apr 2026

- Developed **three PyTorch deep learning models** for fish segmentation and detection under a Ministry of Earth Sciences (MoES)-funded research program, owning the full pipeline: data preprocessing and augmentation (OpenCV, Albumentations), training, hyperparameter tuning and benchmarking.
- Designed **PIAU-Net**, an Attention U-Net with physics-informed constraints and auxiliary supervision, achieving **97.38% mIoU / 98.18% Dice** on the Large Fish dataset and outperforming U-Net, Attention U-Net and DeepLabV3+; reached **93.98% mIoU / 96.85% Dice** after fine-tuning on AquaOV255.
- Built **FishSegDet** (ConvNeXt-V2 + BiFPN), a single model that performs detection and segmentation in one forward pass, reaching **74.34% mAP and 94.41% mIoU** on the USIS16K benchmark, **7.04 points** above YOLOv11l-seg.
- Compressed the approach into **LiteFishSeg** (MobileNetV3 + BiFPN + FCOS) for low-resource edge devices: **9.08M parameters, 2.7× fewer than YOLOv10l-seg**, while retaining **80.30% mIoU and 45.10% mAP**.
- Trained on an NVIDIA RTX 6000 Ada with mixed precision and documented ablation studies and failure analysis, published in the **Indian Journal of Technical Education** (Dec 2025).

## Technical skills

- Programming: Python, JavaScript
- Databases: PostgreSQL, MySQL
- Machine Learning: PyTorch, scikit-learn, deep learning, CNNs, Transformers
- LLMs & RAG: Hugging Face Transformers, PEFT, QLoRA, RAG, Qdrant, vector embeddings
- Computer Vision: OpenCV, TorchVision, Albumentations, U-Net, YOLO
- Deployment & Tools: FastAPI, Docker, Git, GitHub Actions, Linux, MLflow, ONNX

## Research publication

Physics-Informed Attention U-Net (PIAUNet): An Enhanced U-Net Framework for Underwater Segmentation in Aquaculture

Mahboob Alam, S. V. Dhavale, D. Srikanth

Indian Journal of Technical Education, Vol. 48, No. 2 | Dec 2025

Underwater images are degraded by light scattering, turbidity and backscatter, which shifts pixel statistics in ways purely data-driven segmentation models generalize poorly across. PIAU-Net builds the physics into the network: a physics branch at the U-Net bottleneck predicts transmission and backscatter maps, physics-aware attention gates use them to suppress optically unreliable regions in every skip connection, and a composite loss adds a physical reconstruction-consistency term to the segmentation objective.

- [Journal](https://www.ijteonline.in/special_issues)
- [Code](https://github.com/MahboobAlam0/piaunet)
- [Model weights](https://huggingface.co/mahboobalam0/piaunet)

## Distill: Fine-tuning Qwen2.5-3B for Text-to-SQL

Fine-tuned a 3B model with QLoRA on a free Kaggle T4. Exact match on text-to-SQL went from 2.33% to 30.67%.

I wanted to see how far a small open model could get on text-to-SQL with careful fine-tuning and no paid hardware. I fine-tuned Qwen2.5-3B-Instruct with QLoRA on one free T4 and shipped the result as an API and a small web app.

Tools and technologies: QLoRA, Qwen2.5-3B, PEFT, BitsAndBytes, FastAPI, Docker

- [Project write-up](https://alamm.me/project/distill-qwen-text2sql)
- [Source code](https://github.com/MahboobAlam0/Distill---Fine-tuned-Qwen2.5-3B)
- [Demo or model](https://huggingface.co/spaces/mahboobalam0/DistillFinetunedQwen2.5-3B)

## Clinical RAG: Evidence-Based Medical QA

Answers medical questions from 5,355 PubMed abstracts, with citations and a confidence level on every answer.

A question-answering system over PubMed abstracts. It finds the relevant abstracts, answers only from them, cites them, and says how confident it is. If it can't find good evidence, it says so instead of guessing.

Tools and technologies: RAG, Qdrant, LLM (Groq), SentenceTransformers, FastAPI, Docker

- [Project write-up](https://alamm.me/project/clinical-rag)
- [Source code](https://github.com/MahboobAlam0/ai-clinical-rag)
- [Demo or model](https://huggingface.co/spaces/mahboobalam0/ai-clinical-rag)

## HMT-ECGNet: Lightweight ECG Triage

A 0.34M-parameter model that flags abnormal 12-lead ECGs for a doctor to review. It runs on a CPU in about 10 to 30 ms per ECG.

A small model for ECG screening. It reads a 12-lead ECG and gives a probability for each of five PTB-XL diagnostic classes, so the most worrying ECGs can be reviewed first. It's meant to help with triage, not to diagnose.

Tools and technologies: PyTorch, FastAPI, Edge AI, Signal Processing

- [Project write-up](https://alamm.me/project/hmt-ecg-net)
- [Source code](https://github.com/MahboobAlam0/hmt_ecg_healthmonitoringsystem)
- [Demo or model](https://huggingface.co/spaces/mahboobalam0/hmt_ecgnet)

## Physics-Informed Attention U-Net

A U-Net for underwater segmentation that also models how water scatters light. 97.38% mIoU on the Large-Scale Fish dataset, published in a peer-reviewed journal.

This was my thesis research. Underwater images lose contrast and color in ways that depend on the water, and standard segmentation models don't account for that. PIAU-Net adds a small physics branch to an Attention U-Net so the network learns which parts of the image to trust.

Tools and technologies: PyTorch, Computer Vision, Deep Learning, Research

- [Project write-up](https://alamm.me/project/piau-net)
- [Source code](https://github.com/MahboobAlam0/piaunet)
- [Demo or model](https://huggingface.co/mahboobalam0/piaunet)

## FishSegDet: Unified Detection & Semantic Segmentation

One model that draws boxes around fish and segments them in the same pass. 74.34% mAP on USIS16K, 7 points above YOLOv11l-seg.

Counting fish needs boxes, and measuring them needs pixel masks. Usually that means two models. FishSegDet produces both from one network, trained on the USIS16K underwater dataset.

Tools and technologies: PyTorch, ConvNeXtV2, BiFPN, DFL, ASPP, Computer Vision

- [Project write-up](https://alamm.me/project/fish-seg-det)
- [Source code](https://github.com/MahboobAlam0/FishSegDet)

## LiteFishSeg: Edge-Optimized Fish Instance Segmentation

A small version of FishSegDet for edge devices: 9.08M parameters instead of 204.75M, with 80.3% mIoU.

FishSegDet is accurate but far too big for hardware on a boat or a buoy. LiteFishSeg does the same job, boxes plus masks, with a model about 22 times smaller.

Tools and technologies: PyTorch, MobileNetV3, BiFPN, Edge AI, Computer Vision

- [Project write-up](https://alamm.me/project/lite-fish-seg)
- [Source code](https://github.com/MahboobAlam0/LiteFishSeg)

## ResearchKit AI: Private Research Intelligence

A self-hosted research assistant: a paper library with semantic search, chat over your papers, and literature summaries.

I wanted one place to keep papers, search them by meaning and ask questions about them without uploading everything to a cloud service. ResearchKit runs locally, and only the passages needed to answer a question go to the LLM API.

Tools and technologies: FastAPI, Qdrant, LLM (Groq), BGE Embeddings, Docker

- [Project write-up](https://alamm.me/project/research-toolkit)
- [Source code](https://github.com/MahboobAlam0/research-toolkit)
- [Demo or model](https://mahboobalam0.github.io/research-toolkit)

## Fish Density Monitoring System

A web app that counts fish in underwater images and video, maps how crowded each zone is, and raises alerts. It runs on PIAU-Net.

This puts my PIAU-Net model to work. You upload images or video, it segments the fish, splits the frame into zones, counts the fish in each zone and flags zones that get too crowded. It's live on Hugging Face Spaces.

Tools and technologies: PyTorch, PIAU-Net, Gradio, Seg-Grad-CAM, Hugging Face

- [Project write-up](https://alamm.me/project/fish-monitoring-system)
- [Source code](https://github.com/MahboobAlam0/fish-monitoring-system-using-piaunet)
- [Demo or model](https://huggingface.co/spaces/mahboobalam0/fish-density-monitoring-system)

## FinSight: Financial Sentiment & Narrative Intelligence System

Pulls recent news about a company, scores the sentiment with FinBERT, summarizes it, and has an LLM write up the takeaways and risks.

Type a company name and FinSight fetches recent articles, rates each one's sentiment with a finance-specific model, summarizes them, and produces a short report with the main takeaway and risks.

Tools and technologies: Python, FastAPI, FinBERT, LLM (Groq), Docker, Streamlit

- [Project write-up](https://alamm.me/project/finsight-ai)
- [Source code](https://github.com/MahboobAlam0/finsight-ai)

## Urban Mobility Operations Center

An ELT pipeline and dashboard for NYC Citi Bike data that shows which stations are gaining or losing bikes.

A data engineering project on Citi Bike trip data. It loads raw trips, cleans them, builds a star-schema warehouse, and powers a Streamlit dashboard of station inflow and outflow.

Tools and technologies: Python, MySQL, Streamlit, Docker

- [Project write-up](https://alamm.me/project/urban-mobility-ops)
- [Source code](https://github.com/MahboobAlam0/urban-mobility-ops)
- [Demo or model](https://urbanmobilityops.streamlit.app/)

## ArduPilot CI/CD: SITL Automation Pipeline

A CI pipeline that boots the ArduPilot flight software in simulation and runs automated checks on every change.

Every push builds ArduPilot's SITL simulator in Docker, starts a simulated quadcopter and runs automated MAVLink checks against it in GitHub Actions. If a check fails, the change can't be merged.

Tools and technologies: Docker, GitHub Actions, Python, ArduPilot SITL, MAVLink

- [Project write-up](https://alamm.me/project/ardupilot-devops)
- [Source code](https://github.com/MahboobAlam0/ardupilot_devops)

## Learning notes

69 notebooks covering Python Basics, Object-Oriented Python, Advanced Python, Problem Solving, NumPy, Pandas, Data Visualization, Databases, Scraping & ETL, Data Cleaning & EDA, Maths for Machine Learning, Machine Learning, Deep Learning, LLM Systems & Agentic AI, Project Write-ups.

- [Learning index](https://alamm.me/blog)

## Public profiles and contact

- [Portfolio](https://alamm.me/)
- [Contact](https://alamm.me/#contact)
- [GitHub](https://github.com/MahboobAlam0)
- [LinkedIn](https://www.linkedin.com/in/mahboobalam786/)
- [Hugging Face](https://huggingface.co/mahboobalam0)