Harshad Suryawanshi

Transforms property search with natural language — ask in plain English, get semantically matched listings. Combines vector similarity, LLM reasoning, and SQL for hybrid retrieval.

• Vector similarity search via Qdrant for semantic matching
• Text-to-SQL for structured property filters (beds, price, location)
• LlamaIndex orchestration layer for multi-step LLM reasoning

No-code RAG pipeline configurator — test any combination of LLMs, embedding models, and vector stores, then export production-ready Python code in one click.

• Interactive UI to compare LLM + embedding + vector store combinations
• One-click export of production Python code for custom RAG setups
• Supports GPT-3.5/4, Gemini Pro, Cohere, Pinecone, Qdrant, and more

Voice-first LLM + RAG interface designed for accessibility. Say "C3" to activate — no typing required. Built for users who struggle with traditional text-based AI interfaces.

• Wake-word activation ("C3") with real-time voice transcription
• RAG over documents (Nvidia 10-K report) + general LLM queries
• React.js frontend paired with Python Flask + LlamaIndex CreateLlama backend

Zero SQL knowledge required. Query complex databases using plain conversational language — Na2SQL translates your question into accurate SQL and returns interpreted results.

• Natural language → SQL via OpenAI GPT-3.5 + LlamaIndex SQLDatabase
• End-to-end ecommerce database demo included out of the box
• 50+ commits of iterative refinement for edge-case accuracy

Pokémon Go-inspired animal identifier that proves powerful multimodal AI doesn't require GPT-4V. Snap an animal photo — get an AI-generated fact card powered by ResNet + LLM + Wikipedia.

• ResNet18 deep learning vision model for species classification
• Cohere LLM + Wikipedia for rich, contextual animal fact generation
• LlamaIndex orchestrates multi-source retrieval without GPT-4V

A custom GPT-4V alternative built from open models. Upload an image, get an automatic caption from KOSMOS-2, then hold a full conversation about it using PaLM.

• KOSMOS-2 for automatic image captioning via Replicate API
• Google PaLM for contextual multi-turn conversation about images
• Session-based chat history with reset — accessible Streamlit UI

LLM-powered analyst that reads NYSE 10-K filings and generates comprehensive equity research reports. Turns hundreds of pages of dense financial text into actionable insights.

• SubQuestionQueryEngine decomposes complex research into focused sub-queries
• Google PaLM API + bge-small-en embeddings over pre-indexed 10-K documents
• 67 commits — covers 7 major NYSE companies with pre-built vector indexes

Semantic search for Airbnb listings that understands intent, not just keywords. Ask for "a cozy place near the beach for a family of four" and get semantically matched results.

• Qdrant vector database + FastEmbed for lightweight, fast embeddings
• Mixtral 7Bx8 via Groq API for high-speed LLM inference
• LlamaIndex orchestrates retrieval + LLM reasoning pipeline