Headline

GenAI Engineer | RAG | Multi-Agent AI | LLM Evaluation and RLHF | Python | FastAPI

About

I build GenAI systems and multi-agent pipelines with a focus on RAG, LLM evaluation, and autonomous agent workflows. At Turing I have spent the last 2 years on the engineering side of LLM systems. This includes designing adversarial prompts to expose model failure modes across hallucination, instruction-following, and factual accuracy; building a team-level RAG knowledge base for a 400+ person project; running CUA (computer-use agent) evaluations using FastAPI-based verifiers to benchmark agent performance against Gemini and Claude APIs; and training privacy-aware AI assistants using persona-based trajectory data. I also led evaluation squads of 7 to 8 trainers and contributed to SFT and RLHF pipelines. Projects I have built independently: 1. MarketMind: multi-agent marketing system using Google ADK with Research, Strategy, Content, and Critique agents. Live SSE streaming, LLM hot-swap between Ollama and Anthropic, FastAPI backend. 2. AI Research Agent: autonomous BFSI report generator using CrewAI with Researcher, RAG, and Writer agents. Combines Serper web search with ChromaDB and Groq LLaMA 3.3 70B. 3. HerRight: multilingual legal assistant for women in India. Voice input in 7 Indian languages, grounded in official government documents, built with Gemini 2.5 and Sarvam AI. Stack: Python, FastAPI, LangChain, LangGraph, CrewAI, Google ADK, FAISS, ChromaDB, OpenAI, Anthropic, Gemini, Groq, Ollama, HuggingFace, Streamlit, Docker

Location

Bengaluru

AI stack

anthropic apis

cursor

hugging face

openclaw

huggingface

Ragas

Work experience

LLM Python Developer at Turing, Sep 2024 - Present · 2 yrs 1 mo, United States, - Designed 50+ adversarial edge-case prompts per project targeting hallucination, instruction drift, and refusal bypass across 4-5 LLMs simultaneously; executed in sandboxed Docker environments. - Evaluated model outputs across instruction-following, code execution, factual correctness, and knowledge recency. Wrote ideal reference solutions with 100% test pass rates for fine-tuning pipelines. - Coordinated evaluation quality across 7-8 trainers within a 400+ contributor team. Ran daily syncs, reviewed trainer outputs, maintained quality benchmarks, and managed delivery timelines. - Built a team-level RAG knowledge base using LangChain, FAISS, and sentence embeddings for a 400+ person contributor project, allowing trainers to query internal documentation directly instead of searching across multiple docs, with team-scoped access controls isolating retrieval by contributor group. - Worked on CUA (computer-use agent) evaluation: wrote structured UI task prompts covering login, navigation, and assignment workflows, then built FastAPI and Swagger-based verifiers to benchmark agent task completion against Gemini and Claude APIs. - Trained privacy-assistant models using persona-based trajectory data, creating synthetic interaction flows aligned with defined privacy policy constraints. - Contributed to SFT and RLHF dataset pipelines including data curation, response validation, and quality control across coding and NLP evaluation tracks.

System Associate at Infosys, Oct 2022 - Jun 2024 · 1 yr 9 mos, Bengaluru, Karnataka, India, - Automated operational workflows using Python and PowerShell scripts, reducing manual effort across enterprise IT processes. - Used SQL and Python scripts for operational data tasks including incident reporting, log parsing, and data extraction across 10+ internal systems. - Managed 100+ monthly incidents via ServiceNow with 98% SLA compliance; coordinated with cross-functional teams to diagnose and resolve system issues.

Education

Master of Computer Applications - MCA, Data Science, Jain (Deemed-to-be University)

Bachelor of Computer Applications, Information Technology, Amity University Patna

Public links

github: https://github.com/richakumari0311

richakumari.netlify.app: https://richakumari.netlify.app/