AI Developer &
Systems Builder
Specializing in LLMs, RAG Pipelines, and
Agentic AI.
B.Tech CSE @ BIT Mesra (3rd Year)
Key Highlights
What I actually buildLLM Architectures
End-to-end Coding LLMs, Auto-regressive models, and GPT-style SLMs.
Fine-Tuning
Specialized Instruction and Classification fine-tuning for domain specific tasks.
Agentic Workflows
Designing autonomous LangGraph agents and Model Context Protocols (MCP) for complex logic.
Semantic Analysis
Building high-performance Retrieval pipelines powered by FAISS, embeddings, and RAG architecture.
Full-Stack AI
Deploying robust FastAPI and Next.js scalable backends for interactive native applications.
Model Deployment
Pushing end-to-end containerized machine learning endpoints directly into production environments.
Selected Work
All 15 projectsMEVA — Universal Text Classifier
Non-autoregressive classifier fine-tuning Qwen2.5-0.5B with LoRA — 15+ task families resolved in a single forward pass at 8.9× throughput.
MoDenoise — Masked Diffusion LM
A 57.5M-parameter masked diffusion language model generating text by parallel denoising over 128 steps — an alternative to autoregression.
Coding LLM From Scratch
End-to-end LLM built completely from scratch — tokenization through pretraining. Demonstrates architecture-level depth.
By The Numbers
Verifiable, not decorativeTech Stack
Hover to pauseCurrently
Updated 2026MEVA — Universal Text Classifier — a non-autoregressive classifier that fine-tunes Qwen2.5-0.5B with LoRA to resolve 15+ task families in a single forward pass, trading generative decoding for fast, calibrated label prediction. Read the breakdown
5th semester, B.Tech CSE at Birla Institute of Technology, Mesra — holding an 8.5 CGPA alongside independent AI work.
The Dev Log — five long-form entries on persistent LLM memory, reasoning models, RAG optimisation and consumer-hardware fine-tuning. Browse the log
Claude with Amazon Bedrock — Anthropic Academy, 2026. Preceded by OCI 2025 Generative AI Professional, Oracle, 2025.