A timeline, in brief.

• Led RxFact, a benchmark of prescription images testing whether large vision-language models can detect clinically harmful errors; first author (under review, ACL Rolling Review, October 2026).
• Co-led BanglaMedVQA, a physician-validated bilingual medical VQA benchmark, and evaluated open, closed, and medical LVLMs under zero-shot, chain-of-thought, and LoRA fine-tuning (ACL Findings 2026; AAAI Bridge 2026).
• Leading StreamTTRL, a study of continual test-time reinforcement learning with verifiable rewards across languages, domains, and modalities (in preparation for ACL 2027).
• Applied staged PEFT (further pre-training, LP-FT, then LoRA) to low-resource Bangla classification, placing top-15 across all three BLP-2025 Bangla hate speech subtasks.

PyTorch · LoRA / PEFT · LVLM evaluation · Test-time RL · Chain-of-thought prompting · Dataset construction & annotation

• Co-authored AD-SFL (Computer Networks, Elsevier, under review) on activation-space anomaly detection and reinforcement-learning-based defense against data poisoning in SplitFed Learning.
• Present and critically evaluate ML and network-security research papers in regular reading-group sessions.
• Collaborate on ideation, experimental design, and peer review of ongoing research within the group.

PyTorch · Split Learning · Federated Learning · Reinforcement Learning

• Architected IRIS, a multi-agent copilot framework on MCP, owning the orchestration layer that lets independent agents plan, call tools, and hand work off across enterprise systems.
• Built its data backbone on AWS Glue, S3, Lambda, and SageMaker, running enterprise-scale extraction and transformation end to end.
• Deployed and evaluated custom models through AWS Bedrock, standing up reproducible, high-throughput inference for systems in production use.
• Engineered retrieval-augmented generation over RedisVL and pgvector and fine-tuned LLMs in PyTorch, the same adaptation stack behind my published work.
• Profiled and rebuilt FastAPI services with agent-to-agent and WebSocket transport, cutting response time 20% against a measured baseline.

Python · PyTorch · FastAPI · MCP · Multi-agent orchestration · AWS Bedrock · SageMaker · Glue · RAG · PostgreSQL · Redis · WebSockets

• Promoted to a full-time Product Management and Research role to oversee product strategy, feature development, and technical execution.
• Directed a foundational research initiative to create a comprehensive Bangla Speech-to-Text dataset, advancing regional NLP capabilities.
• Led product discovery, feature prioritization, and roadmap planning for AI-driven SaaS products.

Product Strategy · LLMs · OpenAI API · NLP Research

• Engineered "Affordable AI," a pay-as-you-go chatbot service integrating various LLMs, scaling the platform's user base from 1,000 to 20,000.
• Implemented backend services and API integrations to support multi-model LLM routing.
• Collaborated with cross-functional teams to deliver features on schedule.

Python · FastAPI · React.js · OpenAI API · LLMs

• Developed reusable Vue.js components focused on optimal maintainability and developer experience.
• Built backend services using FastAPI, ensuring clean API contracts and fast response times.
• Leveraged TypeScript for robust type safety, enhancing code maintainability and reducing runtime errors.

Vue.js · TypeScript · FastAPI · Python