Projects Experience Education Leadership Skills Beyond Contact Journal NEW
M.S. Applied ML · UMD · Graduating May 2026

Mukul
Rayana

AI/ML Engineer · LLMOps · Game AI · Reinforcement Learning

AI/ML Engineer finishing an M.S. in Applied Machine Learning at the University of Maryland (May 2026). I design and ship production-grade AI — multi-agent systems, LLMOps pipelines, RAG architectures, and reinforcement learning agents that run in real environments. I'm drawn to the hard problems: making AI reliable, fast, and useful where it actually matters.

Explore Projects Resume GitHub ↗ LinkedIn ↗ HuggingFace ↗
Mukul Rayana

Mukul Rayana

About Me
Degree
M.S. Applied ML
University
University of Maryland
Status
Open to Work
Location
College Park
PythonC++C# PyTorchLangChain LangGraphvLLM RAGQLoRA PineconeRay RLlib Unity SentisONNX GuardrailsDocker MLflowFastAPI CUDA

Projects

🎬 gif coming soon
🏆 1st Place · UMD Smith Agentic AI Challenge 2026
LangGraph · XGBoost · Multi-Agent · FastAPI · $4K Prize

PharmaChain — AI Cargo Monitor

Won 1st place and a $4K team prize at the 2026 UMD Smith Agentic AI Challenge. Built a LangGraph orchestration of 8 specialized agents for pharmaceutical cold-chain risk triage. Hybrid risk engine fuses 8 deterministic checks with XGBoost (ROC-AUC 0.9446) and SHAP explanations. RAG-based GDP/FDA compliance checks over 417 regulatory chunks. Human-in-the-loop approval gates — only 2–3 of the 8 agents use an LLM. The rest are deterministic by design.

ROC-AUC 0.94468 agents417 reg. chunksBuilt in 30 days1st Place $4K
Live Demo ↗ GitHub ↗
🎬 gif coming soon
Reinforcement Learning · Game AI · Ray RLlib · Unity Sentis · Groq

Nemesis — Distributed RL Boss AI with Real-Time LLM Narration

PPO boss agent trained with 3-stage curriculum (300K steps, reward −2.4→13.2). Ray RLlib distributed across 8 workers for 2.6× speedup. Exported to ONNX → Unity Sentis for <2 ms/frame inference at 60 fps. Real-time LLM taunts via Groq + NeMo Guardrails (114–352 ms). RL boss outlasts scripted FSM by 85% across 50 episodes.

2.6× RLlib speedup<2 ms/frame60 fps85% outlast FSM
GitHub ↗
🎬 gif coming soon
NLI · Safety Guardrails · Emotion-Conditioned RAG · DeBERTa · captum

EmpathRAG — NLI Safety Guardrail & Emotion-Conditioned RAG

DeBERTa-v3-base fine-tuned on 232K NLI pairs — 0.9629 crisis recall across 30 adversarial probes in 6 attack categories. captum Integrated Gradients on every safety intercept. 5-stage emotion-conditioned RAG over 1.67M FAISS vectors (RoBERTa + LoRA). Ablation: 0.88 emotion alignment vs. 0.30 BM25 — Wilcoxon p = 3.62×10⁻⁸.

0.9629 recall1.67M vectorsp = 3.62e-80.88 alignment
GitHub ↗
🎬 gif coming soon
Multi-Agent · LangGraph · Agentic AI · Staleness Detection

RECON — Agentic ML Research Navigator

4-agent LangGraph state machine with retry loop. Critic applies PASS/STALE/CONTRADICTED/INSUFFICIENT verdicts — catches 52% of superseded ML claims vs. 0% single-pass RAG (130-question eval). Linear decay optimal across 3 ablated formulas. Position accuracy 43.9% vs. 32.3% baseline.

52% staleness catch130-question eval43.9% position acc.
Live Demo ↗ Source ↗
🎬 gif coming soon
LLMOps · QLoRA Fine-Tuning · RAG · CI/CD · HuggingFace Spaces

Irminsul — Production LLMOps System

Llama 3.1 8B fine-tuned with QLoRA (rank 16). Best checkpoint from 3 MLflow experiments: sim 0.826, ROUGE-L 0.466. FastAPI + LangChain + Pinecone RAG with guardrails and confidence-gated web fallback. Auto corpus pipeline ingests 840 docs → 6,876 vectors weekly via GitHub Actions.

Sim 0.826ROUGE-L 0.4666,876 vectors840 docs
Live Demo ↗ Source ↗
🎬 gif coming soon
Document Intelligence · RoBERTa · ONNX · BM25 · BGE · Security

DocPilot — Secure Document Intelligence Platform

Production-hardened QA architecture with prompt-injection defense, PII redaction, RBAC, and append-only audit logging. RoBERTa-squad2 optimized with ONNX INT8 to 90ms P95 latency and 66.0% end-to-end F1. BM25 + BGE dense retrieval with cross-encoder reranking for source-grounded answers.

90ms P9566.0% F1ONNX INT8Prompt Injection Defense
GitHub ↗
🎬 gif coming soon
IEEE Publication · GenAI · VR · Stable Diffusion

Voice-Driven Panoramic VR Generation — IEEE IDCIoT 2024

Speech → 360° image pipeline: Whisper (speech recognition) → GPT-Neo (text generation) → Stable Diffusion (image synthesis). Attention slicing and xFormers optimizations delivered 2.3× throughput within a 10 GB VRAM budget. Peer-reviewed and published at IEEE IDCIoT 2024.

2.3× throughput10 GB VRAMIEEE Published
Read Paper ↗

Experience

Jun 2025 – Present · College Park, MD
Operations Supervisor — Part-Time (Climbing & Bouldering Facility)
Campus Recreation — University of Maryland
  • Promoted from belay staff to supervisor within 7 months; manage daily operations, scheduling, and staff coordination for a 75+ patron/day facility across a 55 ft top-rope wall and 13 ft indoor bouldering zone.
  • Lead shift teams of 4-6 staff, conduct safety briefings, and handle real-time incident assessment, documentation, and emergency response — building the same decision-under-pressure instincts that apply to production incident management.
  • CPR/First Aid certified (American Red Cross). Designed and delivered onboarding training for new hires on safety protocols and equipment inspection procedures.
Jun 2023 – Apr 2024 · Chennai, India
AI Research Intern
BeeBox Studios
  • Built a transformer-based QA system for a 100-page XR corpus achieving 83.2% F1 and 71.8% Exact Match — accurate enough to replace manual in-headset lookup entirely.
  • Cut inference latency 3× (185ms → 62ms P95) via ONNX export and INT8 quantization; shipped via Docker and GitHub Actions CI/CD to AWS EC2.
  • Showcased leadership in technical plan drafting and coordinating a cross-functional team on WebRTC integration for real-time XR data streaming.

Education

Aug 2024 – May 2026 · College Park, MD
M.S. Applied Machine Learning
University of Maryland, College Park
  • Relevant coursework: Deep Learning, Generative AI, NLP, Computer Vision, Cloud Computing, Algorithms
  • Clubs: AI & ML Club · UMD Figure Skating · UMD Archery · UMD Gaming
2020 – 2024 · Chennai, India
B.Tech. in AI and Data Science — GPA 3.55 / 4.0
SRM Easwari Engineering College
  • IEEE Publication: Voice-Driven Panoramic VR Generation (IDCIoT 2024) — speech-to-360° image pipeline with 2.3× throughput improvement.
  • IBM Data Science Professional Certificate (Coursera) · UC San Diego Data Structures & Algorithms (Coursera)

Leadership & Volunteering

May 2023 – May 2024 · Chennai, India
Head of Operations — AI & Data Science Department
SRM Easwari Engineering College
  • Planned and executed symposium events for the Department of AI and Data Science, coordinating 100+ intercollegiate participants across technical and non-technical competitions.
  • Managed logistics, vendor coordination, and cross-team scheduling for multi-day events — building the same operational discipline that later shaped how I plan ML experiment pipelines.
May 2022 – May 2023 · Chennai, India
Head of Photography — Campus Life
SRM Easwari Engineering College
  • Led a team of photographers and videographers covering campus events. Managed end-to-end content production from shoots to final delivery across social media and print.

Skills

Languages

PythonSQLC++C#Bash

GenAI & LLMs

LangChainLangGraphPEFT / LoRA / QLoRARAGAgentic AIMulti-Agent SystemsPrompt EngineeringFine-tuningvLLMPineconeChromaDBGuardrailsRagasGroq

ML & Deep Learning

PyTorchTensorFlowHuggingFace TransformersReinforcement LearningScikit-learnOpenCVNumPyPandas

MLOps & Infra

MLflowFastAPIDockerONNXUnity SentisRay RLlibAWS (EC2/S3)Azure (ACR/ACI)GitHub Actions CI/CDHuggingFace SpacesCUDALinux

Game AI & Engines

Unity 2022.3 LTSUnity ML-AgentsUnity SentisC# ScriptingONNX RuntimeNeMo Guardrails

Beyond Code

Let's Build Something Extraordinary

M.S. Applied Machine Learning · University of Maryland · May 2026
Seeking: AI/ML Engineer · LLMOps · GenAI · Game AI · RL Engineer
STEM OPT eligible · Open to relocate

✉ Email Me LinkedIn ↗
GitHub ↗ HuggingFace ↗ Resume ↗

Or drop me a message directly: