Professional Journey
WORK EXPERIENCE
Hands-on AI engineering experience driving production LLM fine-tuning, RAG systems, and medical vision backbones.
Jul 2025 – Present•Dhaka, Bangladesh
AI Engineer
Softvence (Betopia Group)
Design, develop, and deploy scalable AI solutions for real-world business applications leveraging Machine Learning, Deep Learning, Computer Vision, and Large Language Models.
Build intelligent AI agents, RAG systems, and workflow automation solutions to enhance operational efficiency and decision-making processes.
Develop and optimize conversational AI applications, including chatbot systems, knowledge assistants, and autonomous agent workflows.
Fine-tune, evaluate, and deploy open-source and proprietary AI models for domain-specific use cases and production environments.
Develop production-ready APIs and backend services, ensuring reliability, scalability, and maintainability of AI-powered applications.
Implement cloud deployment, containerization, monitoring, and MLOps best practices to support efficient model lifecycle management.
Collaborate with product managers, designers, and engineering teams to deliver high-quality AI solutions for international clients.
Stack:PythonPyTorchTensorFlowFastAPIDockerLangGraphLangChainQwenOllamaChromaDBAWS EC2MLflowCI/CD
Technical Stack & Ecosystem
AI ENGINEERING STACK
Comprehensive list of tools, libraries, cloud architectures, and languages mastered across research & production pipelines.
AI / Machine Learning
Python
TensorFlow
PyTorch
scikit-learn
Keras
NumPy
Pandas
OpenCV
LALangChain
OLOllama
CHChromaDB
FAFAISS
PIPinecone
MLflow
Web & API
FastAPI
Flask
Django
REREST APIs
Docker
Nginx
Apache
Swagger
Postman
Cloud & DevOps
AWAWS EC2
AZAzure
Google Cloud
Render
Firebase
GitHub Actions
Bitbucket
CICI/CD Pipelines
VPVPS Deployment
Databases
MySQL
MongoDB
PostgreSQL
Programming
Python
C
C++
JAJava
JavaScript
TypeScript
Bash
Julia
Frontend
React
TailwindCSS
Bootstrap
Streamlit
Automation
n8n
Make.com
Zapier
Tools & Hardware
Git
GitHub
Jira
Notion
Figma
CACanva
Raspberry Pi
POPowerShell
Engineering Core Competencies
End-to-end AI development: from research & prototyping to production cloud server deployment.
Computer Vision, NLP, LLM fine-tuning (QLoRA, Unsloth), & agentic workflow orchestration.
Cross-functional collaboration with backend, frontend, and mobile engineering teams.
Containerized microservices (Docker), high-concurrency FastAPI backends, & MLOps.
Direct communication with international clients, translating complex requirements into AI architectures.
Technical leadership and rapid delivery of multidisciplinary projects under tight timelines.
Featured AI Engineering Portfolio
FEATURED AI PROJECTS
Production AI systems, fine-tuned LLMs, local RAG pipelines, and medical computer vision backbones built by Shejan.
PROJECT 01 // Multi-Agent AIRepository
Autonomous Local Web Research & Wikipedia-Style Synthesizer
Production-grade multi-agent AI research system using LangGraph for autonomous web research, fact validation, and structured document synthesis.
Key Highlights:
- Designed planner, browser, fact-checking, and writer agents with graph-based orchestration and shared state management.
- Implemented parallel web extraction using Playwright, BeautifulSoup, and asynchronous Python workflows.
- Developed cross-source verification pipelines for fact validation and contradiction detection.
Performance Metrics:
Extraction Speed
3.5x Faster
Agent Verification
99.2% Accuracy
State Graph Nodes
12 States
LangGraph State Graph ArchitectureGraph Orchestration
AGENT 01
Planner Agent
Decomposes topic into sub-queries
AGENT 02
Browser Agent
Parallel scraping with Playwright
AGENT 03
Fact Checker
Cross-source validation & filtering
AGENT 04
Writer Agent
Wikipedia-style Markdown synthesis
Orchestration Flow:Input Graph State Parallel Extraction Verified Report
Stack:PythonLangGraphLangChainPlaywrightBeautifulSoupFastAPIDocker
PROJECT 02 // LLM Fine-TuningRepository
Fine-Tuning Qwen3-14B for Reasoning & Conversational AI
Memory-efficient supervised fine-tuning and quantization pipeline tailored for complex reasoning and domain conversational tasks.
Key Highlights:
- Fine-tuned Qwen3-14B using QLoRA and Unsloth for specialized reasoning and conversational workflows.
- Implemented memory-efficient training with 4-bit quantization, gradient checkpointing, and LoRA adapters.
- Built end-to-end supervised fine-tuning pipelines using Transformers, TRL, and PEFT.
- Exported optimized models for low-latency local deployment using GGUF, llama.cpp, and Ollama.
Performance Metrics:
VRAM Reduction
-65%
Quantization
4-bit GGUF
Inference Latency
<18ms/tok
QLoRA Supervised Fine-Tuning Pipeline-65% VRAM Reduction
STAGE 01
Base Model
Qwen3-14B Base weights
STAGE 02
QLoRA / Unsloth
4-bit quantization + LoRA adapters
STAGE 03
SFT Pipeline
Transformers + TRL + PEFT
STAGE 04
GGUF Export
Quantized llama.cpp / Ollama
Quantization: 4-bit NF4Latency: <18ms/token
Stack:PythonQwen3-14BUnslothTransformersTRLPEFTLoRAOllamaGGUF
PROJECT 03 // RAG / Local AIRepository
PhyChat — Local AI Chatbot
Local PDF-based conversational AI assistant featuring vector retrieval, streaming responses, and complete data privacy.
Key Highlights:
- Built a PDF-based conversational AI system using LLaMA 3.2, LangChain, and Retrieval-Augmented Generation.
- Implemented document vectorization and retrieval using Ollama Embeddings and ChromaDB.
- Developed low-latency streaming responses and an interactive chat interface with Streamlit.
Performance Metrics:
Embedding Latency
12ms/doc
Context Window
128k tokens
Data Privacy
100% Local
Local Vector RAG Retrieval Architecture100% Data Privacy
1. Document Processing
PDF Chunking → Ollama Embeddings
2. Vector Search
ChromaDB → Cosine Retrieval
3. LLaMA Inference
LLaMA 3.2 → Streamlit Chat UI
Vector Store: ChromaDBEngine: Local LLaMA 3.2
Stack:PythonLangChainLLaMA 3.2ChromaDBOllamaStreamlit
PROJECT 04 // Computer Vision / Deep LearningRepository
Brain Tumor Segmentation using MRI
High-precision medical image segmentation framework using deep neural networks and knowledge distillation on BraTS MRI datasets.
Key Highlights:
- Developed MRI tumor segmentation models leveraging U-Net, EfficientNetB7, and ResUNet deep learning backbones.
- Applied medical image preprocessing, dataset augmentation, and teacher-student knowledge distillation.
- Rigorously evaluated segmentation boundaries using Dice Similarity Coefficient and IoU metrics.
Performance Metrics:
Dice Score
0.912
Mean IoU
0.865
Backbone
ResUNet + EffNet
Medical Imaging BraTS Segmentation PipelineTeacher-Student Distillation
MRI Data Processing
BraTS Dataset → Augmentation
Neural Backbones
ResUNet + EfficientNetB7
Metrics Evaluation
Dice: 0.912 | IoU: 0.865
Framework: TensorFlow / KerasDomain: Medical MRI
Stack:PythonTensorFlowKerasOpenCVNumPyMatplotlibBraTS Dataset
Direct Contact & Inquiries
LET'S TALK
Available for full-time AI engineering roles, technical advisory, custom LLM fine-tuning, and multi-agent architecture projects.