Hi, I’m Kunj Rathod, an AI engineer and researcher working on LLM/RAG systems in healthcare and legal tech.
No. KR-2026.06Salt Lake City — RedmondPortfolio · Preprint · June 2026
Systems that retrieve, reason, & remember.
Kunj Rathod1,2
1 Kahlert School of Computing, University of Utah · 2 Microsoft Fabric, Azure Data (SWE Intern, Jan 2026–)
Abstract
AI engineer and researcher working where production systems meet open research problems. Shipped a HIPAA-compliant LLM platform serving 90+ hospital executives on AWS Bedrock; scaled hybrid legal retrieval to 10M+ documents and 5,000+ daily queries; built graph-augmented pipelines for aerospace materials discovery with NASA and DoD collaborators. Current research targets memory for embodied agents — Video Mind Palace cuts long-horizon QA inference time 31–57% by replacing scene-graph world models with direct video-level VLM queries. Now building distributed data systems for Microsoft Fabric, Azure’s Spark-native lakehouse analytics platform.
Software Engineer Intern — Microsoft Fabric (Azure Data)
Microsoft
—Building distributed data systems for Microsoft Fabric — Microsoft's unified, Spark-native lakehouse analytics platform (the Databricks class of system): OneLake storage, data engineering, warehousing, and real-time intelligence under one compute fabric.
—Full-stack work across the Fabric platform within the Azure Data organization.
Software Development Intern, AI Services (SUDO Program)
University of Utah Health
—Built and deployed a HIPAA-compliant AI chat platform for 90+ hospital executives using React/TypeScript, Flask middleware, and AWS Bedrock microservices with event-driven Lambda orchestration.
—Shipped 6 full-stack features across 4 sprints; integrated AWS Bedrock Agents, Knowledge Bases, and Guardrails for production clinical workflows.
—Reduced inference latency by 40% and data query speed by 60% via Bedrock pipeline optimization, API caching, and a DynamoDB–RDS hybrid database strategy.
—Implemented token-streaming LLM responses (p95 <200ms TTFT) with resilient fallback handling and distributed session persistence for 1,000+ conversations.
—Integrated interactive data visualization tools into the LLM chat interface enabling real-time analytics on hospital data.
—Developed Video Mind Palace (VMP), an efficient agent for Long-term Active Embodied Question Answering (LA-EQA) that replaces scene-graphs with direct video-level VLM queries.
—Demonstrated a 31–57% reduction in online inference time per query with minimal accuracy trade-offs compared to state-of-the-art Robotic Mind Palace (RMP).
—Conducted a comprehensive analysis of the LA-EQA benchmark, identifying key limitations in interactivity and proposing future directions using SceneSmith-generated environments.
STARS Lab, University of Utah · with NASA, Microsoft, U.S. DoD
—Built a multi-agent, graph-augmented pipeline to extract and normalize material-property data from 1,000+ materials-science papers into a physics-aware graph for automated Ashby plot generation.
—Developed a constraint-based 'design region' engine (temperature, creep, pressure limits) and benchmarking suite to identify feasible materials for extreme aerospace environments.
—Built Ref-RAG, a custom RAG chatbot using LangChain and Chainlit to extract structured information from large unorganized PDF datasets for materials researchers.
[Python][LLMs][Multi-Agent][LangChain][Graph RAG]
Nov 2024 – Apr 2025
Remote
AI Engineering Intern
CourtEasy.ai / Nugen
—Scaled hybrid legal-document retrieval to 10M+ indexed Indian legal documents (statutes, court orders), supporting 5,000+ daily queries.
—Improved retrieval accuracy by 28% and reduced hallucinations by 35% via hybrid RAG (dense vectors + BM25 + reranking) and context-grounding optimizations for Legal-NER tasks.
—Built production ETL ingesting 500k+ documents/week and benchmarked 8 LLM families on 4 legal benchmarks including LegalBench and NyayaAnumana; analysis guided model routing, reducing projected inference spend by $50k+/year.
[RAG][BM25][Legal-NER][LegalBench][Python][LLMs]
May – Aug 2024
Salt Lake City, UT
AI Research Intern — BioGraphRAG
Garje Marathi Global (GMG Summer of Code)
—Led development of BioGraphRAG: a Graph Retrieval-Augmented Generation platform combining biomedical knowledge graphs with LLMs for explainable biomedical Q&A.
—Engineered distributed GraphRAG system managing 1M+ biomedical entities (proteins, genes, diseases) integrating UniProt, AlphaFold, and RXNav with NebulaGraph.
—Improved factual accuracy by 40%; optimized graph traversal 3× through strategic caching and high-degree node pruning, achieving sub-500ms query latency at p95.
—Presented at an international AI panel attended by experts from India and the US — received commendation for technical leadership.
—Spearheaded campus-wide outreach driving adoption of Perplexity's AI search among students, faculty, and clubs; onboarded 150+ Perplexity Pro users.
[AI Advocacy][Community][Growth]
Aug – Dec 2024
Salt Lake City, UT
Community Advisor
University of Utah Housing & Residential Education
—Provided conflict mediation, crisis response, and student support services for a 200+ resident community.
[Leadership][Crisis Management]
1.1 Honors
Mar 2026
Kahlert Impact Prize $1,000
Kahlert School of Computing, University of Utah
Awarded for societal impact through AI research and production systems in healthcare, legal-tech, and embodied AI. Funded by a $15M endowment from The Kahlert Foundation; recognizes students translating computing research into real-world benefit.
1.2 Education
Aug 2023 – Dec 2026
B.S. Computer Science
University of Utah · GPA 3.7/4.0, Dean’s List
Machine Learning · Computer Vision · NLP · Distributed Systems · Algorithms & Data Structures
· 2 ·
§2
Selected Work
figures 01–10, in order of recency
Fig. 01 — interactiveResearch
Video Mind Palace
Efficient long-term active embodied QA
—Agent for LA-EQA that replaces expensive scene-graph world models with direct video-level VLM queries.
—31–57% reduction in online inference time per query, within 13–17% of state-of-the-art RMP accuracy.
Live plate — CCD inverse kinematics at 60 Hz. The arm tracks its target autonomously; drag on the floor plane to take over the policy.
Fig. 02Local AI System
FNDR
Privacy-first local AI memory for macOS
—Zero-trust, local-only memory assistant in Rust/Tauri — no cloud, no telemetry, full data sovereignty.
—Metal-accelerated on-device inference (Llama 3.2, SmolVLM) for low-latency RAG on Apple Silicon.
—Real-time screen extraction via Apple Vision OCR + CLIP embeddings; Graphiti-style temporal search across activities, web sessions, and meeting transcripts.
—Local Whisper transcription and an MCP server for secure interop with external agents and IDEs.
—Fully autonomous recruiting backend with specialized agents — Enrichment, Scheduling, Interview, Evaluation — managing the hiring lifecycle from GitHub sourcing to live screening.
—Twilio real-time voice AI interviews, Google Calendar slot scheduling, Slack/Resend approval and outreach flows.
TypeScript · Node.js · PostgreSQL · Anthropic API · Twilio
—Unified intelligence layer over Gmail, Google Calendar, Slack, and FNDR private memory — Smart Todos, natural-language scheduling, on-demand personal context.
—Real-time voice interaction and autonomous multi-step workflow orchestration across the digital stack.
K. Rathod, N. Subedi.“Exploring Long-term Active Embodied Question Answering in Simulated Indoor Environments.”Advanced AI, University of Utah, Spring 2026.
Investigates the LA-EQA benchmark and proposes Video Mind Palace (VMP), an efficient agent replacing structured scene-graphs with direct VLM queries — 31–57% reduction in inference time with minimal accuracy drop, plus a structural critique of benchmark interactivity and episode length.
System architecture, the GraphRAG algorithm, node-degree performance analysis, and a multi-stage answer-enrichment pipeline integrating UniProt, AlphaFold, and RXNav.
✦ NebulaGraph's marketing team requested republication on their official website (Jun 2025).
V2V/V2I/V2N protocols, DSRC and C-V2X standards, real-time speed recommendation algorithms, LSTM traffic-flow prediction, and privacy/security design for V2X systems.
[4]
K. Rathod et al..“Comparative Analysis: LLM Families on Legal Benchmarks.”Internal Technical Report, CourtEasy.ai / Nugen, 2025.
InLegalBERT, InLegalLLaMA, and GPT-4o-mini evaluated on LegalBench and NyayaAnumana, synthesizing 15+ papers to inform production RAG workflow design and model routing.
Fig. 3.2 — The attention motifs this section writes about, run over the site’s own thesis. Synthetic weights, real mechanisms — previous-token, subject-binding & semantic-recall heads. Executable — press ▶ to re-run.
· 4 ·
A
Appendix: Proficiencies
Table 1.Technical proficiencies, grouped by domain.
Domain
Stack
Languages
Python · TypeScript / JavaScript · Java · C++ · Swift · Rust · SQL
Plate A — Proficiency manifold, rendered as a T(2,3) torus knot. Vertices mark sampled skills; topology is, of course, a joke. Interactive — drag to rotate.
· 5 ·
✉
Correspondence
Open to research collaborations and engineering problems worth losing sleep over.