━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
AI Research Engineer and CSE graduate with hands-on production experience building generative AI systems that improve business processes and operational efficiency: RAG pipelines, LLM-based agents, and multimodal AI applications.
I design and ship end-to-end AI pipelines using LangChain, LangGraph, and FastAPI, serving open-source LLMs (DeepSeek, LLaMA, Qwen, Gemma) via Ollama — from schema discovery and prompt engineering through responsible AI practices, production hardening, and deployment. I'm comfortable integrating AI solutions with APIs, databases, and enterprise applications, and I've solved 900+ competitive programming problems along the way.
focus:
generative_ai: [RAG pipelines, LLM agents (LangGraph), fine-tuning (QLoRA/Unsloth), hallucination mitigation]
research: [low-resource NLP, uncertainty quantification, multi-judge evaluation]
engineering: [FastAPI, production hardening, model serving, full-stack delivery]
open_to:
- AI/ML engineering roles
- Applied NLP research collaborations
- Full stack / backend engineering positions━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
|
Languages
Frontend
|
Backend & Databases
AI / ML
|
Systems, Tooling & Deployment
Generative AI & ML: RAG Pipelines · AI Agents (LangGraph) · LLM Fine-Tuning (QLoRA, Unsloth) · Prompt Engineering · Hallucination Mitigation · Text Classification · Evaluation Metrics
Frameworks & Libraries: PyTorch · Transformers · Hugging Face · LangChain · FAISS · ChromaDB · Scikit-learn
APIs & Model Serving: FastAPI · Ollama (DeepSeek, LLaMA, Qwen, Gemma) · Claude API · Gemini API · Whisper STT · F5-TTS
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
| Domain | Proficiency | Details |
|---|---|---|
| Retrieval-Augmented Generation | Advanced | Production RAG pipelines with LangChain, FAISS, and ChromaDB; hybrid search and evaluation design in independent showcase work |
| LLM Agents & Orchestration | Advanced | LangGraph state machines for classification, drafting, self-QA, and human-in-the-loop approval with interrupt() and SqliteSaver checkpointing |
| Hallucination Detection & Mitigation | Advanced | Thesis research on Bangla LLM hallucination detection (UQ + meta-classification, ~96.5% accuracy); production self-correcting feedback loop for Whisper transcription errors |
| LLM Fine-Tuning | Advanced | QLoRA/Unsloth fine-tuning workflows; Qwen2.5-7B-Instruct fine-tuned as a judge model |
| Model Serving & Deployment | Advanced | Ollama-served open-source LLMs (DeepSeek, LLaMA, Qwen, Gemma); FastAPI + Docker deployment on Hugging Face Spaces |
| Speech & Audio ML | Advanced | Whisper STT and F5-TTS integration in production dubbing pipelines |
| Natural Language to Database Queries | Advanced | Custom MCP server + LangGraph enabling local Ollama models to execute natural-language-to-MongoDB queries with zero schema data-bleed |
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
AI Video Dubbing Pipeline — Eagle3D Streaming
An end-to-end, FastAPI-driven AI video dubbing pipeline built in production, combining Whisper transcription with F5-TTS synthesis.
| Stack | FastAPI, Whisper, F5-TTS, FFmpeg, LangGraph, Claude API |
| Reliability | GPU semaphores serializing workloads for 99.9% pipeline uptime under heavy concurrent load; race conditions resolved, silent exception swallowing eliminated, circuit breaker pattern added to reduce synthesis failure cascades |
| Quality | Self-correcting LLM feedback loop that dynamically corrects Whisper hallucinations, cutting manual transcript post-editing time by 40% |
| Cost | Full-stack script-editing and segment re-stitching workflow reducing compute overhead and re-rendering costs by over 80% via targeted per-segment re-synthesis |
| Extras | Automated YouTube SEO feature generating titles, keywords, metadata, and articles from transcriptions via context engineering |
| Repository | [add repo link] |
NL2MQL — Natural Language to MongoDB Query Engine
A custom Model Context Protocol (MCP) server paired with LangGraph, enabling local Ollama models to execute MongoDB queries from natural language.
| Stack | MCP server, LangGraph, Ollama (Gemma 4), MongoDB |
| Architecture | Custom MCP tooling exposing schema-safe query execution to a local LLM |
| Safety | Zero schema data-bleed by design |
| Repository | [add repo link] |
Hallucination Detection in Bangla LLM Output — Thesis
A hybrid framework combining uncertainty quantification with an LLM-as-judge and a meta-classifier to detect hallucinations in LLM-generated Bangla text.
| Stack | LaBSE, XLM-RoBERTa/XNLI, BanglaT5, fine-tuned Qwen2.5-7B-Instruct, XGBoost |
| Dataset | Synthetic Wikipedia-derived and translated HaluEval branches |
| Result | ~96.5% accuracy via UQ features combined with a Platt-scaled meta-classifier |
| Benchmarking | Multi-classifier comparison (XGBoost, TabNet, SAINT, TabPFN) and multi-judge evaluation (Qwen, LLaMA, Mistral, DeepSeek, TigerLLM) |
| Output | Co-authored survey accepted at ICECTE 2026 (IEEE) |
| Repository | [add repo link] |
Customer Support Agent
A support-ticket agent combining retrieval with a stateful reasoning graph for classification, drafting, and self-correction.
| Stack | LangChain, LangGraph, Chroma, FastEmbed, Groq API, Streamlit, SQLite |
| Retrieval | Chroma + FastEmbed embeddings for grounded ticket responses |
| Orchestration | LangGraph state machine handling query classification, drafting, self-QA, and cyclic rewrite-on-failure |
| Human-in-the-loop | High-value refund approvals gated via LangGraph's interrupt(), with SqliteSaver checkpointing for durable pause-and-resume |
| Repository | [add repo link] |
HabiMate — Knowledge-Based Website Chatbot
A retrieval-augmented chatbot delivering context-grounded answers from a custom knowledge base.
| Stack | Streamlit, LangChain, FAISS, FastAPI, Hugging Face, Google Gemini |
| Retrieval | FAISS + Hugging Face embeddings for accurate, source-grounded answers |
| Performance | Google Gemini integration with session caching to reduce inference latency |
| Repository | [add repo link] |
Customer Churn Prediction
An end-to-end ML pipeline predicting telecom customer churn, served in real time and containerized for deployment.
| Stack | XGBoost, Scikit-learn, FastAPI, Docker, Hugging Face Spaces |
| Serving | FastAPI backend for real-time predictions |
| Deployment | Containerized with Docker and deployed on Hugging Face Spaces |
| Repository | [add repo link] |
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
| Role | Company | Period | Focus |
|---|---|---|---|
| AI Research Engineer | Eagle3D Streaming (Remote) | March 2026 – Present | Production AI video dubbing pipeline (Whisper + F5-TTS + FastAPI), NL2MQL, and reliability hardening |
| Tech Solutions Architect | Green Lead (Remote) | 2024 – March 2026 | Climate impact modeling pipelines and full-stack visualization dashboards |
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
| Institution | Program | Period |
|---|---|---|
| Chittagong University of Engineering & Technology | BSc in CSE | 2020 – 2026 |
| Holy Cross College | Higher Secondary Certificate | 2018 – 2020 |
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
| Recognition | Details |
|---|---|
| Champion — NodeScape (CUET Computer Club) | 1st place in a 3-phase flagship hackathon; built an AI-powered graph classification system with algorithm visualization and Dockerized CI/CD deployment |
| Finalist — Intra CUET Machine Learning Contest (IEEE CS CUET) | Built a Bangla physics MCQ question-answering system using LLMs for concept-based reasoning |
| 6th Place — Misogyny Meme Detection (LDK Competition) | Applied multimodal deep learning for automated content moderation in an international competition |
| First Runners-up — ASME Efx Environmental Case Competition | 2nd place internationally, presenting a sustainability solution before a panel of global industry experts and academics |
| Hackathon Judge — CuraAI submission | Mapped evaluation findings to a scoring rubric and flagged inconsistencies between the model card and codebase |
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
building:
- AI video dubbing pipeline (Whisper + F5-TTS + FastAPI) at Eagle3D Streaming
- NL2MQL: natural-language-to-MongoDB query engine via custom MCP server
researching:
- Hallucination detection in Bangla LLM output (thesis, ICECTE 2026)
- Uncertainty quantification for low-resource NLP
exploring:
- Multi-judge LLM evaluation frameworks
- Multimodal AI applications
open_to:
- AI/ML engineering roles
- Applied NLP research collaborations
- Full stack / backend engineering positions━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
