Welcome!

About Me

Hi! I'm Riddhimaan, a research intern at Graphite Growth, an answer engine optimization (AEO) company, where I study whether retrieval-augmented generation degrades once a model's own answers find their way back into the corpus it retrieves from. I also work on the internal platform and the MCP tooling the team uses to support clients.

Last summer I was an AI for Commonwealth intern with the Commonwealth of Massachusetts, where I built a RAG chatbot that made the Unity HPC platform's documentation searchable for more than 500 users.

Outside of work I keep up with new AI research and contribute to open source, including Langfair and Dify. I would rather build something people actually use than something that only looks good in a demo.

💼 Previous Experience

Research Intern

Graphite Growth, Inc. (AEO), Amherst, MA

Feb 2026 – Aug 2026

RAG collapse research: Investigating whether retrieval-augmented generation degrades when its own AI-generated answers re-enter the retrieval corpus, mentored by Graphite’s Chief AI Officer.
GPU evaluation stack: Built a vLLM/LiteLLM serving and evaluation stack with an LLM-as-judge harness, orchestrating 100+ dual-GPU SLURM jobs across a 1,400-question benchmark.
Prompt optimization: Built a GEPA reflective prompt-optimizer with an entity-clustering collapse metric and ran multi-objective Pareto search over anti-collapse vs. answer quality.
Agentic RAG benchmark: Replicated the HotpotQA distractor benchmark and shipped an agentic-RAG variant with a model-driven retrieve() tool, hardened with async concurrency limits, backoff retries, and tool-call fallbacks.
HPC orchestration: Re-architected the job launcher to co-locate inference servers and client in a single GPU allocation, with idempotent, resumable submission that hardened a 108-job sweep against cluster outages.
Semantic search: Shipped pgvector/HNSW prompt matching over SageMaker embeddings, surfacing the top 5 of a 10,000+ prompt bank in 9 ms.
Sampling methodology: Quantified how many ChatGPT responses are needed for reliable brand-visibility estimates across 10,800+ responses, establishing that 10 responses put ~93% of prompts within 10% error and cutting sampling cost up to 5×.
MCP tooling: Exposed platform operations as MCP tools so the assistant reaches parity with the web app, computing the topic-ambiguity signal concurrently for zero added latency.

AI & AWS Intern

AI for Commonwealth of Massachusetts (Mass.Gov), Amherst, MA

Apr 2025 – Sept 2025

RAG chatbot: Built a RAG chatbot (Streamlit, LangChain, AWS Bedrock) for the UMass Unity HPC & AI platform serving 500+ users, presented to the Governor of Massachusetts.
Automated data pipeline: Engineered an event-driven AWS Lambda and EventBridge pipeline refreshing 164+ documents weekly into a Bedrock Knowledge Base.
Model selection: Implemented Anthropic-style contextual retrieval via a custom-chunking Lambda and benchmarked four Bedrock models (Claude, Llama, Amazon Nova) on cost, latency, and accuracy.
CI/CD: Deployed to AWS ECS with 99.99% uptime using GitHub Actions, Docker, and CloudFormation, cutting average support response time by 30%.

For more details you may contact me here.

Full Stack Software Developer

BUILD UMass Amherst, Amherst, MA

Sep 2024 – Dec 2025

Mobile app: Built a cross-platform React Native (Expo) LLM chatbot with offline-first SQLite storage.
Admin dashboard: Shipped a Next.js dashboard with JWT auth on an Express/MongoDB backend for a non-profit serving 100+ users.

AI/ML Intern

LTIMindtree, Chennai, India

Dec 2023 – Jan 2024

Migration benchmarking: Designed benchmark functions analyzing a SQL Server to Microsoft Fabric migration across 1,000+ queries.
LLM automation: Optimized the Azure OpenAI API to convert SQL queries between formats, improving query development time by 25%.