Electronics & Telecommunication Engineering Undergrad
AI/ML Engineer • GPU & Edge AI • LLM Systems • CUDA

I work at the intersection of AI, software systems, and hardware.

My main interests are:

• 🧠 LLMs & Generative AI
• ⚙️ AI/ML Engineering
• 🚀 GPU Computing & CUDA
• 📦 LLM Inference Systems
• 🔧 Edge AI & AI Accelerators
• 🔬 Model Evaluation & Benchmarking
• 🧩 RAG & AI Agents
• 🛠️ Production ML Systems

I enjoy going below the abstraction layer — not only using a model, but understanding how it runs, where it becomes slow, and how the complete system can be made faster and more reliable.

B.E. Electronics & Telecommunication Engineering Jul 2023 – May 2027 • CGPA: 9.125 / 10

Coursework: Data Structures & Algorithms • Object-Oriented Programming • DBMS • Operating Systems

Bengaluru, India • Jul 2026 – Aug 2026

Worked on edge AI accelerator systems, real-time video analytics, model deployment, and multimodal GenAI.

• Stabilized a 16-channel 720p25 real-time video-analytics pipeline on an edge AI accelerator by root-causing IDR-burst packet loss to encoder rate control and tuning in-flight request depth to balance accelerator occupancy and tail latency.
• Root-caused silent output corruption in a shipped runtime library using binary disassembly, tracing it to an undocumented 1200-byte metadata cap.
• Onboarded PyTorch/ONNX models to quantized accelerator binaries using ONNX graph surgery, keeping unsupported operators off the NPU.
• Deployed on-device multimodal GenAI involving VLM + Whisper ASR.
• Filed reproducible issues across five subsystems.

Focus: Edge AI AI Accelerators ONNX Quantization VLMs Whisper Runtime Debugging

Pune, India • Dec 2025 – Jul 2026

Worked on clinical LLMs and production AI workflows for medical applications.

• Fine-tuned clinical LLMs using LoRA/PEFT + Hugging Face Transformers for AI-assisted radiology report generation.
• Defined task-specific evaluation metrics for model evaluation.
• Automated medical-document processing including PDF extraction, cleaning, labeling, and annotation into training-ready datasets.
• Prototyped and shipped workflows as FastAPI + MongoDB REST services.
• Worked on Responsible AI systems with monitoring, bias checks, and drift checks.

Focus: LLM Fine-tuning LoRA/PEFT Transformers FastAPI MongoDB Responsible AI

A 4B fine-tuned LLM that outperformed a 70B API model on a task-specific NLU benchmark.

• Fine-tuned Gemma 3 4B using LoRA/PEFT.
• Performs multilingual intent classification and slot filling.
• Supports Bengali, Hindi, Kannada, Malayalam, Tamil, and Telugu.
• Ran training-size ablations using 3K, 9K, and 18K examples.
• Performed leave-one-language-out experiments to study cross-lingual generalization.
• Outperformed a 70B API baseline on the evaluated NLU benchmark.
• Built reproducible training and evaluation workflows.

🔗 GitHub

An LLM inference stack built from the GPU kernel upward.

• Implemented a hand-written CUDA FlashAttention kernel with shared-memory tiling, online softmax, fused attention, prefill, and split-K decode.
• Packaged the kernel as a pip-installable PyTorch extension.
• Served Qwen2.5-0.5B behind an OpenAI-compatible SSE server.
• Added static batching for inference serving.
• Profiled the complete system against vLLM.
• Identified that the major decode bottleneck was host dispatch rather than the CUDA kernel.
• Fixed launch-geometry starvation and recovered approximately 3× decode throughput.
• Validated numerical correctness with 2.44e-4 error against FP32 across 99 tests.

🔗 GitHub

A persistent AI assistant that can manage and edit its own memory.

• Built a MemGPT-style assistant that persists memory across restarts.
• Designed three memory tiers: Core • Recall • Archival.
• Added provenance tracking for stored facts.
• Implemented forced eviction at 95% context usage.
• Achieved 115/115 on a deterministic benchmark across both storage backends.
• Enforced the benchmark in CI on every push.
• Built a four-provider LLM failover router with a real HTTP error taxonomy.
• Protected MCP tools using path isolation and human-approval interrupts.
• Tested prompt-injection defenses against a predefined injection corpus.

🔗 GitHub

A distributed RAG system that turns long-form video into a searchable, timestamp-grounded knowledge base.

Video
  │
  ▼
Whisper Transcription
  │
  ▼
Semantic Chunking
  │
  ▼
FAISS + BM25 Hybrid Retrieval
  │
  ▼
CrossEncoder Re-ranking
  │
  ▼
Timestamp-Grounded Answer

• Built a distributed 6-microservice RAG system.
• Combined dense FAISS retrieval with BM25 lexical retrieval.
• Added CrossEncoder re-ranking.
• Generated answers grounded to video timestamps.
• Used llama.cpp / Phi-3 for CPU inference.
• Added Redis caching.
• Fully Dockerized the system.
• Built a React frontend.

🔗 GitHub

Python C++ C SQL CUDA Bash

PyTorch Hugging Face Transformers LoRA / PEFT LLM Fine-tuning RAG Hybrid Retrieval CrossEncoder Re-ranking AI Agents LangGraph MCP Whisper YOLOv8 VLMs Model Evaluation Benchmarking

CUDA Kernel Development GPU Profiling ONNX Graph Surgery Quantization vLLM llama.cpp AI Accelerators Edge AI

FastAPI REST APIs Microservices PostgreSQL pgvector MongoDB Redis FAISS ChromaDB

Docker Git Linux AWS GitHub Actions CI/CD pytest

• 🎓 AI Agents Course — Hugging Face
• 🤖 Machine Learning Specialization — DeepLearning.AI / Andrew Ng
• 📊 Complete Data Science, ML, DL & NLP Bootcamp — Krish Naik

        AI / ML
           │
           ▼
     Efficient LLMs
           │
           ▼
    GPU & CUDA Systems
           │
           ▼
     AI Accelerators
           │
           ▼
        Edge AI
           │
           ▼
   Real-World AI Systems

I'm particularly interested in the complete path from:

Model → Graph → Runtime → Kernel → Accelerator → Production System

• 🏍️ Bike Riding
• 📚 Reading
• 🥾 Hiking
• 🏏 Cricket
• 🏸 Badminton