Arav Adikesh
Ramakrishnan
>> MS Computer Science @ UMass Amherst
>> Fullstack Engineer & ML Researcher
Graduate student at UMass Amherst (May 2026). Currently at UMass CDS, where I started as a SWE intern building MCP servers for 5,000+ users and transitioned into ML research on electrochemistry potentials. Also a research extern at AI2 evaluating literature grounding for their autodiscovery system.
>> SYSTEM SPECS
● KERNEL
MS Computer Science at UMass Amherst (4.0 GPA, Bay State Scholar), concentrating in Data Science. Graduating May 2026.
Currently a research extern at AI2, building literature grounding for automated scientific discovery, and an ML engineer at UMass CDSAI, developing ML interatomic potentials for electrochemical simulations and fullstack GenAI platform infrastructure serving 5,000+ users.
Prior work spans healthcare NLP (EMNLP 2025), digital forensics, RAG systems, and production fullstack systems — from Spring Boot microservices and Kubernetes deployments to React frontends and MCP server architecture.
● LOADED MODULES
>> EDUCATION RECORDS
Master of Science in Computer Science
University of Massachusetts Amherst
Expected: May 2026 | GPA: 4.0
// Concentration:
Data Science
// Achievements:
- →Graduate Bay State Scholarship
- →Data Science for the Common Good Fellow (Summer 2025)
// Coursework:
Bachelor of Science in Computer Science
University of Massachusetts Amherst
Graduated: May 2024 | GPA: 3.93
// Minor:
Economics
// Achievements:
- →Chancellor's Scholarship
- →Dean's List (All Semesters)
// Coursework:
>> EXECUTION LOG
// Work Experience
Machine Learning Engineering Intern
UMass Center for Data Science and AI
Boston, MA
- →Reproduced the TRECI workflow (transfer learning + DEAL active learning) for generating ML interatomic potentials for Au/water electrochemical interfaces, heavily adapting the active learning framework to work with custom lab data on our university HPC environment.
- →Compiled and optimized GPU-accelerated LAMMPS (Kokkos/CUDA, A100) with MACE pair style; ran comprehensive benchmarking including energy drift, throughput (ns/day), and peak GPU memory profiling to validate simulation stability and performance.
Software Engineering Intern
UMass Center for Data Science and AI
Boston, MA
- →Architected and deployed MCP Document Server enabling multi-format document generation (PDF, DOCX, CSV) directly through LLM chat interfaces for 5,000+ active campus users, with JWT authentication and comprehensive file management, improving content workflow efficiency by ~60%.
- →Built end-to-end file management system with in-browser preview modal, user/group-based storage quotas, and Google Drive OAuth2 integration for seamless import/export, serving as core infrastructure for the campus GenAI platform.
- →Implemented an MCP connector for Amazon Athena, enabling natural language querying of AWS databases through LLMs, cutting query formulation and debugging time by ~75%.
- →Led development of Media Cloud classifier pipeline, a fully automated, containerized BERT-based classifier processing 100K+ news articles from a 2B+ corpus with 96% accuracy.
- →Implemented Optuna-based hyperparameter optimization and dashboard-driven evaluation, boosting reproducibility and deployment-ready ML workflows.
Machine Learning Intern
Prime Focus Technologies
Los Angeles, CA
- →Developed a RAG-powered support chatbot using LangChain and FAISS vector database to handle 500+ daily queries, achieving 88% user satisfaction.
- →Created an automated query classification system that reduced manual triage and saved $15K annually in support costs.
- →Deployed end-to-end production-grade conversational AI systems with JavaScript frontend, Spring Boot microservices, Flask APIs on AWS Lambda, and Kubernetes, achieving <200ms response time and 99.5% uptime.
Undergraduate Course Assistant
UMass Amherst
Amherst, MA
- →Conducted 5+ weekly office hours, assisting 50+ students with code troubleshooting and learning support.
- →Graded 200+ assignments with precision, offering constructive feedback to support student growth.
// Research Experience
Research Extern
Allen Institute for AI (AI2)
Remote
- →Developed and evaluated a literature grounding subsystem for AutoDiscovery, an MCTS-based scientific hypothesis generation system, testing retrieval pipelines across research hypotheses using Semantic Scholar and Google Scholar APIs with LLM-powered query expansion, cross-encoder re-ranking, and async relevance scoring.
- →Designed and tested 4 query expansion strategies (semantic, sub-hypotheses, decomposition, baseline) optimized for Semantic Scholar's LightGBM ranker; implemented async LLM scoring (GPT models) with SHA-256 deduplication, checkpoint recovery, and structured Pydantic outputs.
- →Experimented with cross-encoder (ms-marco-MiniLM) and RankGPT sliding-window re-rankers, and built a novel LGSG metric evaluating how retrieved papers shift LLM belief distributions.
- →Engineered crash-resumable SLURM array job pipeline with atomic writes, exponential backoff, startup jitter, and disk-backed SHA-256 caching for API rate-limit handling.
- →Developed MedCOD framework integrating UMLS and LLM-KB knowledge sources to enhance English-to-Spanish medical translation — improving translation quality by 80% (BLEU ↑ from 24.47 → 44.23) through structured prompting and LoRA fine-tuning.
- →Published research in EMNLP 2025 Findings, contributing a novel approach to domain-specific translation addressing healthcare communication barriers for limited English proficiency populations.
- →Contributed to RescueBox, an open-source digital forensics platform developed by UMass Rescue Lab for processing large-scale digital evidence.
- →Engineered and deployed modular forensic analysis plugins, including deepfake detection and perceptual hash-based image similarity systems.
- →Optimized inference and data pipelines by converting PyTorch models to ONNX, achieving 3× faster inference speeds.
>> PROJECT ARCHIVE
Agentic Disambiguation for Ambiguous Question Answering
Research project investigating agentic RAG (Retrieval-Augmented Generation) for handling ambiguous open-domain questions. Implements LangGraph-based multi-agent system with HyDE (Hypothetical Document Embeddings), sub-query decomposition, and structured multi-interpretation synthesis. Achieved 8.6% F1 improvement over vanilla RAG on AmbigNQ dataset using hybrid retrieval (BM25 + FAISS).
// Dependencies:
Web Agent Security Research
Research project investigating security vulnerabilities and attack vectors in LLM-powered web agents. Analyzes potential risks and proposes mitigation strategies for autonomous agents interacting with web environments.
// Dependencies:
UMass Outing Club Gear Locker
Scalable REST API using Express.js/TypeScript with Firebase Real-time Database, handling 100+ daily transactions. Led a team of 3 developers implementing Agile methodologies.
// Dependencies:
