Selected Projects
Production-grade AI systems with proven performance improvements and cost savings
LatentMAS-SLoRA: Multi-Agent Reasoning with S-LoRA
Multi-agent framework enabling collaboration in latent space rather than token space, augmented with role-specialized LoRA adapters. Featured as 1 of 5 community extensions in the official Gen-Verse LatentMAS repository (766+ stars).
Technical Approach
- Latent Collaboration: Agents communicate via hidden states, reducing token usage by 50-80% and achieving 3-7× speedup
- S-LoRA Integration: Four specialized adapters (Medical, Reward, Comics, Detection) with dynamic hot-swapping
- Multimodal Support: Qwen2.5-VL-7B-Instruct foundation with vision-language reasoning
- Production Infrastructure: RunPod serverless deployment with Docker, CI/CD, and 200ms latency per agent
- Advanced RAG: Document injection via URL, base64, JSON with domain-aware routing
AI Calling Agent Platform
Real-time voice conversation platform with SIP/WebRTC telephony integration. Achieved sub-500ms latency with emotionally expressive speech synthesis.
Multi-GPU Video Generation
Distributed inference pipeline for state-of-the-art video generation (text-to-video, image-to-video, speech-to-video). Achieved 3× throughput increase using FSDP on serverless GPU clusters.
Custom LoRA Training Pipeline
Self-hosted Flux.1 Dev with custom LoRA fine-tuning infrastructure. Delivered 80% cost reduction for client photography workflows.
Enterprise RAG System
Multi-modal retrieval-augmented generation with vector search and semantic chunking. Achieved 40% accuracy improvement over baseline implementation.
Voice-Pro: Speech Processing Platform
Web application for speech recognition, translation, and voice cloning across 100+ languages. Supports YouTube processing and real-time translation.
Medical Imaging with Transformers
Brain tumor classification and segmentation using ConvNeXt V2 and SegFormer. Achieved 99.6% diagnostic accuracy on evaluation dataset.
Professional Experience
Building production AI systems and conducting applied research
AI & Machine Learning Engineer
Freelance — Multiple Clients
2023 — Present
- Develop and deploy cutting-edge ML/AI models specializing in multi-modal tasks including image generation, video synthesis, NLP, and voice AI
- Design and implement serverless GPU infrastructure with Docker and Kubernetes, achieving 60%+ cost reduction
- Build production RAG systems and multi-agent frameworks with measurable performance improvements
Research Assistant
Rajshahi University — Solar Lab / AI Lab
Mar 2022 — May 2023
- Conducted research on renewable energy (solar cells) and speech processing using ML/DL techniques
- Applied machine learning to analyze simulation data and improve photovoltaic performance
- Published 4 peer-reviewed papers in Q1 journals with impact factors up to 7.1
Skills & Technologies
Full-stack ML engineering with production-grade tools and frameworks
AI & Machine Learning
MLOps & Cloud
Frameworks & Tools
Serverless GPU
Languages
LLMs & Models
Publications
7 peer-reviewed publications • 4 Q1 journals (IF up to 7.1) • Google Scholar Profile
Machine learning assisted revelation of the best performing single hetero-junction thermophotovoltaic cell
Sustainable Energy Technologies and Assessments, 2025
Machine Learning-Enabled Performance Exploration of AuCuSe₄ in Thermophotovoltaic Cell
Solar Energy, 2024
Unleashing the Power of Open-Source Transformers in Medical Imaging
Int'l Journal of Advanced Computer Science & Applications, 2024
Numerical studies on a ternary AgInTe₂ chalcopyrite thin film solar cell
Heliyon (Cell Press), 2023
Numerical prediction on the photovoltaic performance of CZTS-based thin film solar cell
Nano Select, 2023
Spectrum estimation for voiced speech using average weighted linear prediction
2024
Enhancement of Bone Conducted Speech Using Deep Transfer Learning
2024
Get in Touch
Open to AI/ML Engineering roles, MLOps consulting, and collaborative research projects