I build backend systems and AI agents, and I make sure they actually work before I call them done.
- 🤖 Agentic AI systems that reason over real-time, with multimodal capabilities
- 🔍 RAG pipelines and context engineering
- ⚡ Production backends (FastAPI, Docker, Cloud Run) built end to end, from API to deployment
I like owning things end to end, from API design to cloud deployment, and I'd rather prove a system works than assume it does. The projects below show that in practice.
A live interview coach that watches you through your camera and gives honest, real-time feedback, on your answers, posture, and eye contact, the moment the session ends.
Orchestrates two concurrent YOLO models (pose and object detection) alongside Gemini's Realtime Video LLM, synchronizing all three within a single live session. Designed and shipped solo in 4 days.
Python YOLO Gemini Realtime
A question-answering system that reads documents and answers questions about them, but checks its own confidence first and admits when it doesn't know something, instead of guessing.
Built with FastAPI, LangGraph, and Pinecone. Grades its own retrieval confidence and does a deeper read on borderline cases rather than trusting a similarity score alone. Validated with an automated 15-case evaluation suite (15/15 pass, including adversarial "should refuse" cases).
Python FastAPI LangGraph Pinecone
A tool that reads an entire codebase and turns it into a clear, structured map, so getting help from an AI assistant on a bug takes one attempt instead of three.
Uses Python's AST module for static analysis, with a custom metric I designed (Context Leverage Factor) benchmarking a 92.3% reduction in representation size. Cuts debugging token usage from ~1,800 to 380 in head-to-head testing, powered by Groq's Llama 3.3 70B.
Python AST Groq API
cross_lens — Adversarial RAG System
A research assistant that checks a claim from multiple angles at once, actively looking for evidence that supports it, contradicts it, or is simply neutral, instead of just confirming what you want to hear.
Implements Corrective RAG with adaptive query routing and multi-vector retrieval. Deployed live on Streamlit Cloud using local embeddings at zero infrastructure cost.
Python Streamlit RAG
Lets someone ask a plain-English business question and get a real answer pulled straight from a company's data warehouse, no SQL required.
Built with Google's Agent Development Kit and Gemini, using a Model Context Protocol (MCP) server for governed, read-only access to BigQuery. Containerized and deployed on Google Cloud Run.
Python Google ADK MCP Cloud Run
An agent that analyzes a coffee shop's sales and scheduling data to spot staffing and inventory issues, then asks for explicit approval before making any changes.
Cross-references point-of-sale data with a schedule to identify operational bottlenecks, runs its analysis in a sandboxed environment, and requires human permission before writing any updates.
Python Google ADK Gemini Cloud Run
- coffee-barista-agent — AI assistant recommending personalized coffee orders using RAG, deployed on Cloud Run.
- Linkedin-Post-Generator — A tool that generates LinkedIn post drafts using an LLM based on a given topic.
- Bengaluru House Price Prediction — A regression model predicting Bengaluru house prices from property features like location, square footage, and BHK count, using scikit-learn.
Languages: Python, SQL, TypeScript GenAI & Agentic AI: LangChain, LangGraph, RAG, AI Agents, Prompt Engineering, Model Context Protocol (MCP) LLM Providers: Gemini, OpenAI, Groq Backend & Cloud: FastAPI, Docker, Google Cloud Run, PostgreSQL Fundamentals: Data Structures & Algorithms, Object-Oriented Programming, DBMS
- Agentic AI Foundations Associate, Oracle
- Large Language Models Course, HuggingFace
- B.Tech in Computer Science and Engineering, Sri Venkateswara University
Currently open to Backend Engineering, Software Engineering, and Applied AI / GenAI roles.
📫 yamini26284@gmail.com Linkedin : https://www.linkedin.com/in/yamini26284/