Intelligence in the interface.
Intention in the architecture.
Retrieval supplies relevant context to the language model. Prompt design connects the question to the evidence.
Turning unstructured information into structured reports with language models and retrieval.
Structured information and reports for document analysis and compliance review.
I’m building BioMedAI around a simple ambition: make complex medical information easier to explore, understand, and work with. From the details of a patient encounter to the extraordinary architecture of the human body, this is where my curiosity about medicine meets my love of building software.
Prescriba brings prescription drafting, patient visits, and trends into one connected prototype. Designed around the workflows of Bangladeshi doctors, it explores how AI assistance, explicit checking rules, and thoughtful interfaces can fit together.
Discover BioMedAI (opens in a new tab)The Human Atlas turns anatomy into an interactive landscape. Explore the body in 3D, search for structures, isolate a part, and look through layers to discover how its systems fit together. Built on the open Z-Anatomy project.
Explore the Human Atlas (opens in a new tab)“The most interesting part is turning complexity into something someone can actually use.”
PERSONAL PROJECT · IN DEVELOPMENTThe problem, the architecture,
and the decisions that connect them.
LLM-Based Document Intelligence System
Make long, unstructured documents useful for analysis and review.
Connect document extraction, relevant context, and report generation in one inspectable workflow.
Structured information and reports for document analysis and compliance review.
RAG Pipeline with Vector Databases
Make information discoverable even when a question and its source use different words.
Combine document chunking, embeddings, vector databases, and hybrid search to assemble useful context.
Semantic retrieval and context-aware responses grounded in the indexed collection.
Computer Vision Video Analysis Pipeline
Turn a stream of images into observations that are easier to inspect and use.
Connect video processing and custom detection models with frame annotation in an OpenCV pipeline.
Annotated frames and object detections for reviewing video content.
Secure FastAPI Backend for AI Services
Make AI services accessible through APIs with explicit boundaries and predictable responses.
Bring authentication, request validation, logging, and error handling into the service layer.
Python backends for AI microservices, with GCP deployment and monitoring.
Agent-Based AI Workflow Automation
Reduce repetitive workflow steps while keeping people involved in important decisions.
Use LangFlow and custom agents to connect tool calls with validation checkpoints.
Controlled AI workflows that support review and human-in-the-loop validation.
OWASP LLM Security Implementation
Reduce the ways untrusted input and generated output can affect connected AI services.
Combine input validation, output filtering, prompt-injection mitigation, and secure API design.
Application protections that address LLM risks across the input, processing, and output boundaries.
Start with the person and the problem. Find the constraints before choosing the technology.
Design the smallest useful system. Keep its boundaries explicit and its decisions easy to follow.
Validate inputs, account for failure, and document the decisions that the next engineer will need.