MIMANSHI SHARMA

I bring hands-on experience in building generative AI systems using RAG and LLMs. Skilled in developing multi-agent pipelines, evaluating model behaviour, and translating research ideas into working POC using modern AI tools.

Multi-Agent Bias Audit Pipeline- 

• A 4-agent AI system that detects and eliminates the bias in LLM outputs via self-correction.
• Pipeline: Generator (Gemma) → Auditor (LLaMA 3.1, ethics policy-driven) → Refiner (Mistral Nemo) → Verifier (LLaMA 3.1)
• Detects gender, cultural, racial, and socioeconomic bias. Inspired by 'Multi-Agent Bias Mitigation via Self-Correction Frameworks' (IEEE, 2024/2025)

Memory AI - Persistent AI Assistant - 

• Built a POC for a RAG-based AI assistant with persistent memory, enabling contextual retrieval across sessions., Implemented embedding-based retrieval for context-aware response generation.
• Deployed scalable system using FastAPI, Docker, and Google Cloud Run.
• Live Demo: https://memory-ai-671546906215.us-central1.run.app/

LLM Peer Review System- 

• Automated generate-then-critique pipeline to improve LLM response quality using local models via Ollama.
• LLaMA 3 generates a response and LLaMA 3.1 critiques it - identifying errors and gaps., Results exported to structured CSV.

LLM as a Judge 

• A multi-model evaluation system where two LLMs debate a topic and a third acts as an impartial judge., Gemma argues for, Mistral argues against across N rounds
• Qwen evaluates the full transcript and outputs with reasoning, confidence scoring, and comparative analysis of responses.

Conversational Arguments with LLMs- 

• Adversarial debate system where two LLMs argue opposing sides across multiple rounds.
• Gemma and Mistral are assigned opposing positions and respond to each other iteratively. Captures and stored full interaction logs for analysis of model behaviour and consistency.