AI Engineer with industry experience building and deploying LLM, RAG, and machine-learning systems. Experienced across PyTorch, retrieval systems, Docker/CI/CD, and production AI deployments, with research focused on transformer robustness and backdoor defense.
01 Work experience
AI Engineer at Data Solution-360
Jan 2026 to PresentDhaka, Bangladesh
Worked on Linx360.com, an AI-powered interview and assessment platform that identifies skill gaps, delivers role-aligned interview practice, and helps candidates become job-ready.
Built a Python/FastAPI MCQ assessment pipeline with OpenAI API, generating 15 personalized questions from candidate profiles and skill gaps; parallelized requests with Redis TTL storage, cutting latency from about 10 to 15 s down to about 2 to 5 s.
Built an LLM critic/regeneration loop with strict JSON schemas and Bloom-aligned rubrics, regenerating rejected questions until all 15 passed validation.
Engineered a scikit-learn ranking/scoring workflow over about 30K records using Pandas/NumPy for feature engineering, leakage-safe proxy targets, and held-out evaluation with Precision@50.
Migrated Docker services from Linux to DigitalOcean App Platform with automated GitHub Actions and GHCR CI/CD.
Used Claude Code and Codex with MCP, Superpowers, and pytest TDD to plan, implement, debug, and validate features end to end.
Python
FastAPI
OpenAI API
Pandas
NumPy
scikit-learn
Redis
Docker
GitHub Actions
GHCR
DigitalOcean
pytest
Claude Code
OpenAI Codex
MCP
Superpowers
02 Projects
Data-Aware RAG System for Research Papers
Local multi-paper research assistant with hybrid retrieval and source/page citations.
Built Python PDF ingestion with Marker and pdfplumber, using page-aware chunking to preserve source and page metadata.
Generated local all-MiniLM-L6-v2 embeddings with Sentence Transformers and combined dense cosine search with BM25 using normalized weighted fusion.
Decomposed multi-paper questions into atomic needs, then rewrote and routed queries to retrieve evidence from relevant papers.
Generated citation-aware answers locally with Ollama, propagating retrieved paper and page metadata into responses.
Parallelized LLM query classification, reducing an eight-query experiment from 70.6 s to 20.3 s with the same 26/35 score.
ClientManager Pro: AI-Powered Client Management System
Full-stack client platform with project workflows, real-time communication, and AI-powered repository Q&A.
Built Next.js frontend and Express/Node.js REST APIs with MongoDB persistence for client and project workflows.
Implemented JWT authentication and role-based access control across multiple user roles and protected routes.
Built GitHub repository Q&A with Transformers.js and all-MiniLM-L6-v2, storing embeddings in Pinecone and answering retrieved-context questions with Gemini.
Reduced repeated AI computation with embedding/answer caching and query deduplication, avoiding unnecessary embedding and generation calls.
This is a summary of my CV. Figures such as latency and speed-up are as reported there; the case studies on this site say what was measured and what was not.
