AI Researcher at Wipro Innovation Labs, working on modern AI systems and advanced AI use cases — agentic pipelines, retrieval-augmented reasoning, and LLM-driven tooling. B.Tech from IIT Madras, with in-depth knowledge spanning machine learning and full-stack engineering, from model behavior down to production infrastructure.

I like taking on complex, ambiguous problems and shaping them into working systems. If you've got an interesting idea, a hard problem, or just want to talk shop — I'm always up for the conversation.

An open-source, AI-powered application using Agentic CAG to chat with any public GitHub repository or developer profile, offering deep code analysis, visual architecture maps, and security audits.

Key Highlights:

• Context-Aware Engine (CAG): Intelligently selects relevant code snippets instead of loading entire repositories
• Visual Architecture Maps: Auto-generates Mermaid flowcharts from complex code logic for instant visualization
• Deep Profile Intelligence: Analyzes coding styles, commit patterns, and developer habits across entire portfolios
• Zero-Config Security Audits: Detects vulnerabilities with AI-powered triage and actionable fix recommendations
• Mobile-First Design: The only tool in its class optimized for on-the-go code reviews

Why It Stands Out:

• Works instantly on any public repo — no installation, login, or GitHub App required
• Uses Context Augmented Generation (CAG) vs traditional RAG for superior code understanding
• Generates interactive visuals instead of text walls
• Analyzes developers, not just code

Tech Stack: Next.js • Google Gemini • Vercel KV • TypeScript • Tailwind CSS • Framer Motion

A voice-first to-do list application for intuitive task management entirely through natural language commands, powered by Deepgram for real-time transcription and Groq's qwen-2.5-32b for AI command analysis.

Key Highlights:

• Voice-First Interface: Create, update, and manage tasks using natural language — no typing required
• Real-Time Transcription: Leverages Deepgram's blazing-fast speech-to-text for instant feedback
• AI-Powered Command Analysis: Understands complex intents like "remind me to buy groceries tomorrow at 5pm"
• Client-Side Intelligence: Local models for instant priority detection and date parsing

Impact:

• 95%+ transcription accuracy with sub-second latency
• Handles 20+ natural language command variations
• Fully responsive with touch and voice support

Tech Stack: Next.js • React • Deepgram • Groq (qwen-2.5-32b) • ShadCN/UI • Tailwind CSS • Framer Motion

An advanced web application providing expert-level road safety intervention recommendations using Retrieval-Augmented Generation (RAG) architecture with Google's Gemini LLM and Vertex AI Search.

Key Highlights:

• RAG Architecture: Grounded recommendations from a curated knowledge base of road safety research
• AI Orchestration: Leverages Google Genkit to manage the entire AI workflow seamlessly
• Evidence-Based Advice: Provides detailed rationale with direct citations from authoritative source documents
• Optimized Query Flow: Brainstorms multiple search queries and executes parallel searches for reduced latency

Impact:

• Processes 100+ authoritative safety documents
• 60% faster response time through parallel search optimization
• Provides source-backed recommendations with verifiable citations

Tech Stack: Next.js • React • TypeScript • Tailwind CSS • ShadCN UI • Google Genkit • Gemini 2.5 Flash • Vertex AI Search • Firebase App Hosting

An AI-powered system for extracting cancer-related information from patient Electronic Health Record (EHR) notes, focusing on information retrieval and structured medical data extraction.

Key Highlights:

• Multi-Stage Information Retrieval: Combines keyword search (BM25) with semantic search (Sentence Transformers, CrossEncoder)
• LLM-Based Data Extraction: Uses quantized Qwen/Qwen2.5-7B-Instruct-1M for structured JSON output
• GPU Efficiency: Utilizes 4-bit quantization for running 7B parameter models on T4 GPUs

Impact:

• Processes complex medical notes with 90%+ extraction accuracy
• Runs efficiently on consumer GPUs through quantization

Tech Stack: Python • NLTK • BM25 • Sentence Transformers • CrossEncoder • Qwen LLM • bitsandbytes • PyTorch

An LLM-powered tool that enables users to extract information, transcribe, and ask questions about YouTube video content, providing a seamless way to interact with video transcripts.

Key Highlights:

• Speech-to-Text: Utilizes OpenAI Whisper for high-quality, multilingual audio transcription
• NLP-Driven Querying: Leverages Hugging Face Transformers for natural language understanding and query resolution

Impact:

• Supports 50+ languages through Whisper
• Enables semantic search across video content

Tech Stack: Python • OpenAI Whisper • Hugging Face Transformers • FFMPEG • NLP

I believe in building AI systems that are:

• Purpose-Driven — every line of code should solve a real problem
• Intelligently Designed — leverage AI where it adds genuine value, not as a buzzword
• Performance-First — optimize for speed and efficiency without sacrificing functionality
• User-Centric — complex technology should feel simple and intuitive
• Open & Collaborative — knowledge grows when shared

"The best AI is invisible — it just works."

I'm always up for talking through an interesting idea, a hard technical problem, or a project worth building.

_{Built by Sameer Verma · Last Updated: August 2026}