Projects

A web-based interactive tool for extracting geographic coordinates from scanned maps and generating precise Ground Control Points (GCPs) and GeoJSON data for automated geo-referencing in GIS applications. Built at BIDS (Advisor: Dr. Atonu Rabbani) — cut per-map processing from a week to 84 maps in weeks and applied the pipeline with NLP-based analysis to assess regional economic impact across Bangladesh.

Hybrid — Apr ‘26 – Jun ‘26 · BIDS Research Assistant.

pdforge is a fast, privacy-first PDF toolkit running on Cloudflare Workers (Edge) — merge, split, compress, rotate, and watermark PDFs without uploading to a central server. Built for clean, modern UX with instant client-side processing and edge caching.

• Live: pdforge.ahmedfahad04.workers.dev
• Edge-native, no backend storage · Clean modern UI

Developed a full-featured B2B marketplace connecting farmers, suppliers, and buyers in the Bangladeshi agriculture sector, incorporating real-world customer feedback for enhanced usability.

Implemented modules for product listing, bidding, payment verification, and order tracking, prioritizing intuitive design and functionality following user requirements.

Designed and implemented a full-stack e-commerce platform with dedicated workflows for both users and administrators.

Developed an admin dashboard supporting inventory management, order processing, and product analytics.

Integrated Google Analytics (via MCP server) to track user behavior and inform data-driven feature improvements.

Served 25+ real users and optimized the platform for mobile-first performance, addressing 94% mobile traffic.

Built a VSCode extension for real-time detection and repair of Python static type errors using LLM-based reasoning.

Implemented context-aware fix generation using Retrieval-Augmented Generation (RAG) over indexed codebases with vector embeddings.

Designed a relational database schema (SQLite) to store code chunks, error metadata, fixes, and user feedback for continuous improvement.

Developed an interactive desktop learning application tailored for autistic children to support cognitive, behavioral, and social development.

Implemented features including task scheduling, gaze-based interaction, and progress monitoring.

Intent-based semantic search using fine-tuned BERT + Qdrant vector DB, deployed on Dockerized EC2 with Grafana & Prometheus monitoring. Improved relevance via language detection, negative-intent handling, LLM-powered filtering, and caching.

2025 — Ranked 6th of 85 teams.

Developed a solution using YOLO, DETR, and Deep Learning models for object detection in the BadODD dataset.

Utilized different road objects from nine districts of Bangladesh to cover a wide variety of objects, ensuring 43% accuracy in vehicle detection.

Processed a 79+ hour Bengali speech corpus to improve transcription accuracy for Bengali dialects.

Designed an end-to-end speech recognition pipeline using deep learning models that achieved 75% accuracy for speeches in specific dialects.

Enhanced Qwen-7B for Bangla reasoning by fine-tuning with curated math datasets.

Applied Direct Preference Optimization (DPO) and used mergeKit to improve inference quality for Bangla math problems.

An NLP-based system using DistilBERT for symptom analysis and specialist mapping.

Provided essential clinic details, including locations, working hours, and appointment scheduling, streamlining the patient-doctor connection.