Projects

Built the Text-to-Speech (TTS) and Automatic Speech Recognition (ASR) modules and pipeline behind REVE's AI Voice Agent, a conversational voice assistant for call centers, for low-resource Bengali with support for regional dialects, using Python, PyTorch, Docker, and HuggingFace — dataset preparation, model finetuning, pronunciation fixes, quantization, converting models to C++ (.so) for production deployment, and Real-Time Factor (RTF) optimization for low-latency performance.

Built the admin and customer-facing portals of REVE's telecom billing and DID (Direct Inward Dialing) numbering platform, using Java, Spring Boot, Vue.js 3, MySQL, Redis, JPA/Hibernate, and AWS S3 — DID lifecycle management (buyer API, multi-provider integrations, an automated buy/renew/hold/release/backorder scheduler), payment orchestration, fraud detection, reporting, HTTP rate limiting, and JWT/RBAC-secured REST APIs for client onboarding and service provisioning.

Distributed balance management library synchronizing financial data between MySQL and Redis in real-time for telecom billing.

Key Responsibilities: Built a chunk-based lease system reducing Redis round-trips by ~70%, with crash recovery and Redis Pub/Sub sync.

Reporting and analytics system for telecom VoIP/DID operations — automated Revenue Assurance and ERP reporting from CDR and inventory data, plus a multi-tenant BI layer with telecom traffic KPIs (Answer Seizure Ratio, ACD, NER, PDD — distinct from the speech-recognition "ASR" elsewhere on this site).

Key Responsibilities: Built 14+ daily RA report types, automated monthly invoicing, multi-tenant BI reporting with flexible aggregation by country/operator/time, and HTTP rate limiting for API protection.

Automated CI/CD pipeline for a full-stack enterprise application (Vue.js 3 + Spring Boot 3), enabling one-click production releases.

Key Responsibilities: Built a multi-stage Jenkins pipeline with zero-downtime deployment, automated rollback, and build notifications.

Microfinance solution supporting Grameen Bank's core operations for disbursing and managing small-scale financial services.

Key Responsibilities: Built microloan processing, savings account operations, and client portfolio tracking.

Stack: ASP.NET, Microsoft SQL Server

Human Resource Management solution for Grameen Bank, streamlining HR operations and employee lifecycle management.

Key Responsibilities: Built employee data management, payroll processing, and leave/attendance workflows.

Stack: ASP.NET Core, ASP.NET Core MVC, SQL Server

Workflow design project for organizational note flow management, covering product requisition, approvals, and digital signatures.

Key Responsibilities: Designed the end-to-end workflow architecture and multi-tier approval hierarchy.

Implemented and evaluated a congestion-counter-based enhancement to the AODV routing protocol for MANETs, improving throughput and packet delivery ratio in denser networks.

Implemented a full CNN (convolution, max-pooling, ReLU, fully-connected, softmax; forward/backward propagation) from first principles using only NumPy, for the BUET Machine Learning Sessional.

Built an embedded patient-queue and vitals-collection system on the ATmega32 microcontroller, with a registration station (temperature and pulse sensors, keypad, LCD) and a doctor's consultation station, simulated in Proteus. Published as an arXiv preprint and submitted to NSysS 2026.

Full-stack travel planning and management web application built with a 3-person team, including a companion service for local transport guidelines.

Implemented memory management (paging and page-fault handling) in the xv6 teaching operating system, for the BUET Operating Systems Sessional.

Implemented AES and RSA from scratch, a buffer-overflow exploit, a Morris-worm-style self-propagating program, and firewall rule design, for the BUET Computer Security Sessional.

Open-source contributions spanning Speech AI — Text-to-Speech (TTS) and Automatic Speech Recognition (ASR) — and Java libraries, plus finetuned models and datasets published on HuggingFace.

Highlights: LoRA finetuning contribution to k2-fsa/OmniVoice, a TTS benchmarking playground (including Real-Time Factor and latency comparisons), finetuned Bengali TTS models — including adding Bengali support to CosyVoice3, which didn't have it by default — with a public live demo, and published datasets/quantized models with several thousand+ downloads.

A RESTful API for a blogging platform built with Spring Boot, with JWT authentication and role-based access control.

Key Features: CRUD for posts/comments, category and tag management, pagination, and image uploads.

A Django-based online marketplace/classifieds platform with product listings, job postings, chat, and search.