Open to AI/ML roles

Mukul Rayana

AI Engineer · Forward Deployed Engineer · agents, RAG & LLMOps

M.S. in Applied Machine Learning, University of Maryland (May 2026). I build agent and retrieval systems that real people use.

Portrait of Mukul Rayana under cherry blossoms
  • 1st place

    UMD Smith Agentic AI Challenge 2026

    PharmaChain, an 8-agent LangGraph system for pharma cold-chain risk

    $4K prize · Team Synapse

  • 3× faster

    Production QA inference at BeeBox Studios

    P95 latency after ONNX export and INT8 quantization

  • IEEE

    Published at IDCIoT 2024

    Voice-driven panoramic VR generation: speech to 360° scenes

    Read the paper

  • 3 live demos

    Agents and RAG you can try right now

    Hosted on Vercel and Hugging Face Spaces

Projects

Eight builds across agents, reinforcement learning, AI safety, research tooling, LLMOps, security, generative AI and civic tech. Each leads with its headline result.

Multi-agent 1st place · $4K prize

PharmaChain: AI Cargo Monitor

UMD Smith Agentic AI Challenge · April 2026 · Team Synapse (5)

12–15 sfrom alert to an approved action plan, versus the usual 6–8 hours
  • Specialised agents (LangGraph)8
  • GDP/FDA regulatory RAG417 chunks
  • Team XGBoost risk model (ROC-AUC)0.9446
  • Deterministic rules + veto8 + 1

Pharmaceutical cold-chain risk triage. A hybrid risk engine combines 8 deterministic rules and a deterministic veto with an XGBoost model, and checks compliance against GDP/FDA guidance through RAG. Only 2–3 of the 8 agents use an LLM; the rest are deterministic by design, and a human approves every action. My part: the Route and Triage agents, their REST API, and failover across four LLM providers with a rule-based fallback.

Architecture
  1. Live shipment telemetryTemperature, shock and routing events stream in
  2. LangGraph orchestrator · 8 agentsRisk scoring runs in three lanes: 8 rules + deterministic vetoXGBoost + SHAPGDP/FDA RAG · 417 chunks
  3. Human-in-the-loop approvalAn operator signs off before anything is acted on
  4. Coordinated actionBackup facilities, carrier reroutes, rescheduled appointments, insurance claims

alert → action in 12–15 s

Cited RAG · LLMOpsLive beta · v1

Irminsul

Genshin Impact Q&A assistant that cites every answer and updates itself after each patch

95.1%live test set (117/123), all 24 leak, injection and safety probes handled
  • Players so far15
  • Knowledge base, rebuilt daily~6,800 docs
  • p95 at 50 users0.45 s
  • Running cost$0

Stats, talents, damage maths and material costs come straight from game data with no AI. Everything else is answered from cited sources, and any sentence they don't support is removed before you see it. It rebuilds itself from game data and the community wiki, a publish gate keeps unreleased content out, and a UID lookup plans what to farm today. Unofficial fan project; theorycrafting from KeqingMains, credited per answer.

A question answered straight from game data, with its source shown. Genshin Impact content belongs to HoYoverse.

Older build (v0.9)

March 2026: a QLoRA fine-tune of Llama 3.1 8B with Pinecone RAG.

  • Semantic similarity0.826
  • Pinecone index6,876 vectors
View on GitHub

hybrid search + re-rankergame data or cited AIgrounding check

Agentic research

RECON

Multi-agent ML research navigator that catches stale claims

44% vs 0%of stale claims caught, versus a single-pass RAG baseline
  • False positives2% (v1: 8%)
  • Benchmark130 questions
  • Agents4 · LangGraph

A critic agent flags research claims that newer work has replaced, so answers point to current evidence instead of outdated papers.

PlannerRetrieverCriticSynthesizer

Reinforcement learning

Nemesis

Distributed RL boss AI with real-time LLM narration

+85%longer survival for the RL boss than a scripted FSM, across 50 episodes
  • PPO curriculum, 300K stepsreward −2.4 → 13.2
  • Ray RLlib, 1 → 8 workers2.6× speedup
  • Unity Sentis inference<2 ms/frame · 60 fps
  • LLM taunts (Groq)<500 ms

PPO boss agent trained with a 3-stage curriculum, exported to ONNX and run in Unity Sentis. Real-time taunts come from Groq behind NeMo Guardrails.

Curriculum PPORLlib ×8ONNXUnity Sentis

Generative AI IEEE IDCIoT 2024

Voice-Driven Panoramic VR Generation

Peer-reviewed speech → 360° image pipeline

2.3×throughput from attention slicing and xFormers
  • VRAM budget10 GB
  • PublicationIEEE IDCIoT 2024

WhisperGPT-NeoStable Diffusion360° panorama

Secure document AI

DocPilot

Secure document Q&A service

4× smallermodel with ONNX INT8 quantization (496.5 MB → 125 MB), 2–3× faster per inference on CPU
  • SearchBM25 + embeddings, reranked
  • AccessRBAC + audit log
  • UploadsPII redaction · injection checks
  • Tests42 unit · 20 smoke

A FastAPI service with role-based access, audit logging, PII redaction on upload and prompt-injection screening. Hybrid search (BM25 + embeddings) with cross-encoder reranking keeps answers tied to their sources, and a web UI sits on top.

AI safetyResearch prototype

EmpathRAG

Safety guardrail for an emotion-aware RAG assistant

0.9629crisis recall on held-out test data; adversarial prompts are still a weak spot
  • Emotion alignment vs BM250.88 vs 0.30
  • FAISS index1.67M vectors
  • Generationlocal Mistral 7B

A DeBERTa-v3 NLI model screens messages for crisis signals, with Integrated Gradients explaining each intercept. Retrieval and generation run fully locally.

NLI safety gateemotion-conditioned retrievalMistral 7B

Civic document AIGDG AI for Good 2026: Housing

Promise Ledger

Turns a housing deal's public records into one checked receipt of its affordable-home promises

811 pagesof real county records: all 21 Prince George's County, MD PILOT agreements since 2023
  • Held-out precision / recall95.7% / 93.6% on 47 claims
  • Planted attacks held8 of 8
  • Blank template text foundin 6 approved deals
  • Model cost to read all 811 pages$0.61

When a county gives a housing developer a tax break, the developer promises affordable homes in return. Promise Ledger reads the deal's public records and lists every promise with the exact quote and page that proves it. The final build adds a night 3D map with a column of light at each address, and "Ask the county", which answers one question across every deal: Gemini writes, code checks every number and date, and each sentence links to its source.

PythonFastAPIGeminiCloud RunFirestore

Experience

Apr 2023 – Apr 2024Chennai, India

AI Research Intern

BeeBox Studios

  • Built a transformer-based QA system for a 100-page XR corpus, reaching 83.2% F1 and 71.8% Exact Match: accurate enough to replace manual in-headset lookup entirely.
  • Cut inference latency 3× (185 → 62 ms P95) with ONNX export and INT8 quantization; shipped with Docker and GitHub Actions CI/CD to AWS EC2.
  • As the only ML engineer on the team, supported six production releases, and delivered multi-peer WebRTC streaming a day ahead of a four-day client deadline.
83.2% F171.8% Exact Match on XR corpus QA
185 → 62 msP95 latency via ONNX + INT8
Shipped withDockerGitHub ActionsAWS EC2

Jun 2025 – May 2026College Park, MD

Climbing Wall Supervisor

UMD Campus Recreation · Climbing Wall Staff, Jun 2025 · Supervisor, Nov 2025 – May 2026

  • Promoted from climbing wall staff to supervisor within 6 months; led shift teams of 4–6 for a 75+ patron/day facility, ran safety briefings and incident response, and designed staff onboarding. CPR/First Aid certified.

Education

Aug 2024 – May 2026College Park, MD

M.S. Applied Machine Learning

University of Maryland, College Park

  • Coursework: Deep Learning, Generative AI, NLP, Computer Vision, Cloud Computing, Algorithms
  • Clubs: AI & ML Club · UMD Figure Skating · UMD Archery · UMD Gaming

2020 – 2024Chennai, India

B.Tech. AI and Data Science

SRM Easwari Engineering College

  • IEEE publication (co-author): Voice-Driven Panoramic VR Generation (IDCIoT 2024)
  • IBM Data Science Professional Certificate (Coursera)
  • UC San Diego Data Structures & Algorithms (Coursera)

Leadership & volunteering

May 2023 – May 2024Chennai, India

Head of Operations, AI & Data Science Dept.

SRM Easwari Engineering College

  • Planned and ran department symposiums for 100+ intercollegiate participants across technical and non-technical competitions.
  • Managed logistics, vendors and cross-team scheduling for multi-day events: the same operational discipline I now bring to ML experiment pipelines.

May 2022 – May 2023Chennai, India

Head of Photography, Campus Life

SRM Easwari Engineering College

  • Led a team of photographers and videographers covering campus events, owning production from shoot to final delivery across social media and print.

Skills

Languages

  • Python
  • SQL
  • C#

GenAI & LLMs

  • LangChain
  • LangGraph
  • PEFT / LoRA / QLoRA
  • RAG
  • Agentic AI
  • Multi-Agent Systems
  • Prompt Engineering
  • Fine-tuning
  • Llama
  • vLLM
  • Groq
  • FAISS
  • Pinecone
  • pgvector
  • ChromaDB
  • Guardrails

ML & Deep Learning

  • PyTorch
  • TensorFlow
  • HuggingFace Transformers
  • Reinforcement Learning
  • Ray RLlib
  • Scikit-learn
  • XGBoost
  • OpenCV
  • NumPy
  • Pandas
  • CUDA

MLOps & Infra

  • MLflow
  • FastAPI
  • Pydantic
  • Docker
  • ONNX
  • AWS (EC2/S3)
  • GitHub Actions CI/CD
  • Supabase
  • HuggingFace Spaces

AI Safety & Eval

  • NLI Guardrails
  • LLM-as-a-Judge
  • Ragas
  • SHAP
  • Captum
  • NeMo Guardrails

Game AI & Engines

  • Unity 2022.3 LTS
  • Unity ML-Agents
  • Unity Sentis
  • C# Scripting
  • ONNX Runtime

Beyond Code

  • Genshin Impact video game

    AR 60 since 2021. Full 36-star Spiral Abyss clears, often with off-meta comps just to see if it's possible. Three years of elemental reaction math and boss pattern memorization. Nemesis? Yeah, that's basically a love letter to Azhdaha.

    My Genshin guides on YouTube (opens in a new tab)
  • Anime

    Shonen battle arcs, psychological thrillers, and anything with a hard magic system. Currently on a hunter exam binge. Always looking for the next arc that hits like Chimera Ant or Shibuya. Send recs.

  • Rock Climbing

    I supervised at UMD's climbing facility and climb regularly. Managing patron safety and sending problems on a 55 ft wall in the same shift was pretty great honestly.

  • Track & Field

    Sprinter: hurdles and 4×100 relay. I ran track competitively and the start-gun adrenaline honestly never goes away.

  • Archery

    Range sessions at UMD Archery Club. Recurve bow, target shooting. Nothing clears the head like a clean release.

  • Figure Skating

    UMD Figure Skating Club. Still learning edges and crossovers but the ice is genuinely addictive.

  • Guitar

    I play acoustic covers when I need a break. Nothing fancy, just vibing with whatever sounds right.

Let’s build something extraordinary

M.S. Applied Machine Learning, University of Maryland, May 2026. Looking for AI Engineer and Forward Deployed Engineer roles, including agent, RAG and LLMOps work. Open to relocate.

0 / 2000

Messages go straight to my inbox. I usually reply within a day or two.