Projects

Systems I have built

Selected systems demonstrating my experience translating machine-learning and software ideas into working applications.

Selected projects

From problem to working system

Personal

Full-stack AI

Problem: Enable shoppers to search an e-commerce catalogue using text, an image, or both while keeping results aligned with the changing product database.
My Role: I built the FastAPI retrieval service and the PostgreSQL-to-vector synchronisation pipeline, with Redis caching and containerised deployment.
Approach: OpenAI vision capabilities interpret image queries, ChromaDB supports hybrid text/image retrieval, PostgreSQL stores catalogue data, and Redis caches repeated requests.
Technologies: FastAPI, OpenAI Vision, ChromaDB, PostgreSQL, Redis, Docker.
Result: A working full-stack search and recommendation system with synchronised relational and vector data.
Resources: GitHub repository

Personal

Computer vision

Problem: Estimate cattle weight from an ordinary photograph when a livestock scale is unavailable.
My Role: I developed the end-to-end computer-vision and regression pipeline and served it through a FastAPI backend.
Approach: YOLOv8 instance segmentation isolates the animal and a reference object; the reference provides image scale, and regression maps extracted morphometry to predicted weight.
Technologies: YOLOv8, instance segmentation, regression, FastAPI.
Result: A working image-to-weight prediction pipeline designed for photographs from an uncalibrated camera.

Professional

BRAC Tech &
SELP-RHRN
2025

Problem: Support a Bengali-language platform addressing gender-discrimination mindsets through interactive and research-informed content.
My Role: I built the FastAPI backend, multi-LLM content pipelines, and PostgreSQL JSONB schemas.
Approach: Containerised backend services used GPT-4o and Gemini for Bengali book and research-paper search and suggestions, with semi-structured content stored in PostgreSQL.
Technologies: FastAPI, GPT-4o, Gemini, PostgreSQL JSONB, Docker.
Result: A deployed backend developed during the BRAC Tech & SELP-RHRN AI Fellowship.
Resources: Private professional project - source unavailable.

Champion

AI Meets Climate
Hackathon

Problem: Turn spoken field reports into structured climate-incident records quickly enough to support real-time reporting.
My Role: Developed the pipeline as part of the winning team during the hackathon.
Approach: Automatic speech recognition converts audio to text, and named-entity recognition extracts structured incident information.
Technologies: ASR, named-entity recognition, data pipelines.
Result: Champion, AI Meets Climate Hackathon, Oxfam.