Reverse proxy and review layer for monitoring LLM agent systems without SDKs or code instrumentation.

• Architected a Rust-based reverse proxy that captures AI agent traffic transparently without modifying existing codebases.
• Designed an LLM-assisted review layer that generates critiques, suggested labels, failure signals, and confidence scores while keeping human approval as the final decision.
• Added a human-in-the-loop pipeline that converts agent traces into training datasets for RLHF and AI agent evaluation.
• Built with Rust, TypeScript, React, PostgreSQL, Vite, Docker, Kubernetes, and Prometheus.

Brain-inspired AI system that learns from conversations by building semantic knowledge graphs, reducing hallucination versus naive RAG.

• Architected a 4-layer memory system spanning sensory, semantic, episodic, and working memory; built an automated sleep cycle that consolidates memories, prunes noise, and reinforces high-confidence relationships nightly.
• Implemented hybrid graph plus vector RAG using Neo4j, ChromaDB, and Groq inference; extracts clean semantic nodes instead of raw text chunks to reduce token usage and improve reasoning accuracy.
• Created interactive 3D visualizations of agent cognition in real time and deployed the full-stack system on Hugging Face Spaces with Docker, concurrent-user handling, and persistent Neo4j storage.
• Built with FastAPI, React, Neo4j, Groq (Llama 3.1), LangChain, ChromaDB, and Docker.

High-throughput job intelligence pipeline that extracts skills, knowledge, and tasks into a graph.

• Designed and shipped a data ingestion pipeline that processes 2M job postings/day to extract and structure skills, knowledge, and tasks.
• Scaled LLM inference to 800K unique jobs/day using vLLM with async batch workers on serverless GPU cloud.
• Built with Python, vLLM, Llama 3.1, FAISS, Neo4j, Kafka, and Modal.

Open-source LLM pipeline for turning resumes, syllabi, and job postings into structured KST data.

• Parses unstructured text into Knowledge, Skills, and Tasks for downstream skill intelligence workflows.
• Serves 5+ organizations in production as part of LAiSER research and deployment work.
• Uses PyTorch, Hugging Face Transformers, Sentence-Transformers, FAISS, Neo4j, vLLM, and GitLab CI/CD.

1st place geospatial child vulnerability system for Cyclone Amphan risk response.

• Built a dashboard quantifying 494K+ children at risk and 666 critical facilities across Bangladesh.
• Ran the analysis purely on geospatial data to model vulnerability and infrastructure exposure around Cyclone Amphan.
• Used Python, GeoPandas, Rasterio, and Plotly to deliver an interactive tool for UN officials.

1st place LLM-powered metadata automation system for open-source GitHub projects.

• Built an automation workflow using Python, LangChain, and the GitHub API.
• Generated useful repository metadata to reduce manual project maintenance work.