RESUME

Vetriselvan R

AI / ML Engineer · Researcher · 2nd-year CS + AI

Delhi NCR, India

Summary

2nd-year CS+AI undergrad working at the intersection of RL and systems — GRPO for compiler codegen, agentic LLM workflows, and spatio-temporal forecasting. 2nd on private leaderboard (Top 4 offline) at IIIT-H × IIT Delhi × IBM AISEHack 2026, with the trained model published to HuggingFace and Kaggle. Meta PyTorch OpenEnv Hackathon Finalist.

Education

B.Tech — Computer Science and Artificial Intelligence

Newton School of Technology, Rishihood University · Focus: Deep Learning, Reinforcement Learning, Agentic Systems · CGPA: 8.1 / 10

Aug 2024 – May 2028 (expected)

Delhi NCR, India

Projects

ARM-Gym — RL Compiler Optimizer

2025 — Present · Core Contributor (3-person team) · Live at kaori02-arm-gym.hf.space

PythonGRPOQwen2.5-Coder-7BLoRAHuggingFace TRLLLVM-MCAQEMUDocker

›GRPO + LoRA fine-tune (V11): Qwen2.5-Coder-7B with LoRA (r=32), 8 generations per step, 250 steps in 107 min on a single L40S; assembly correctness lifted 19% → 70%, mean reward 3.30 → 6.50.
›Three-stage verifier: GNU assembler for syntax, QEMU with 20 adversarial inputs per kernel for correctness, LLVM-MCA (Neoverse V2 model) for cycle-throughput rewards. Owned PR #1 rebuilding the correctness gate to compare stdout against a gcc-compiled reference binary, closing a reward-hacking loophole in the prior link-and-exit check.
›Curriculum and infrastructure: 649 kernel variants across 15 compute templates (GEMM, softmax, dot product, activation) graded by difficulty; AArch64 cross-toolchain Docker image and HF-hosted training logs / eval plots.
›Result: Best measured speedup of +14.5% over clang-21 -O3 on LLVM-MCA estimates.

SRE Incident Responder

Mar 2026 · Meta PyTorch OpenEnv Hackathon × SST · Finalist

PythonFastAPIPydantic v2RedisPostgreSQLDockeropenenv-coreWebSocketMCP

›OpenEnv-compatible RL environment that spins up a real microservice cluster (live Redis and PostgreSQL, not mocks) and challenges agents to debug production outages end-to-end.
›Environment: Five escalating incident tiers from OOM-killed restarts to identifying leaking key patterns amid deliberate red herrings; dense reward shaping (+0.2 diagnosis, +0.8 correct resolution) with penalties for destructive actions.
›Live state grading: Validators query actual infrastructure state instead of mocked traces; multi-protocol access via REST, WebSocket, and MCP (JSON-RPC) for cross-framework agents.

AFNO V4 — PM2.5 Spatio-Temporal Forecasting

Apr 2026 · Team Lead · IIIT-H × IIT Delhi × IBM AISEHack (Track 2) · 2nd on private LB / Top 4 offline

PyTorchAFNOSimVPU-NetEMA EnsembleHuggingFaceKaggle

›Adaptive Fourier Neural Operator (~7M params) forecasting gridded PM2.5 16 hours ahead over India from 10 hours of history on a 140 × 124 grid at ~25 km resolution.
›Architecture: U-Net encoder → multiscale AFNO (1 spectral block at 35 × 31 + 6 blocks at 18 × 16 bottleneck) → SimVP temporal translator (10 → 16 frames) → U-Net decoder predicting a residual on top of persistence.
›Physics-informed inputs: 20 channels including ventilation coefficient, advection tendency, and diurnal-climatology anomaly; two-seed EMA ensemble with 16 per-horizon multiplicative calibration factors on a leak-free validation split.
›Validation-leak fix: Diagnosed overlapping-window leak in the original random split; contiguous month-end holdout with a gap made evaluation honest and unlocked the ~0.86 plateau.
›Result: 2nd on private leaderboard (0.8833); Top 4 in offline Phase 2 judging. Weights and code published on HuggingFace and Kaggle.

Agentic Credit Risk Scoring — LangGraph + RAG

Oct — Dec 2025

PythonLangGraphLangChainChromaDBGroq Llama 3.3 70Bscikit-learnStreamlit

›Cost-sensitive ML: 100+ behavioral features; class weighting lifted Medium-Risk recall from 0.41 to 0.65 at 80.4% accuracy on a 10,268-sample test set.
›Agentic RAG: LangGraph workflow (risk-scoring → retrieval → report generation) over ChromaDB; Llama 3.3 70B (Groq) grounds each recommendation in retrieved regulatory text.
›Live demo: crs.streamlit.app

Skills

Languages

PythonTypeScriptC++SQL

ML & DL

PyTorchTensorFlowscikit-learnNumPyPandasTransformersHuggingFace TRLLoRA

RL & LLM Tooling

GRPOPPOLangGraphLangChainRAGChromaDBOpenEnvMCPLLVM-MCAQEMU

Infrastructure

FastAPIStreamlitDockerGitGitHub ActionsMongoDBNext.jsReact

Certifications

›

Machine Learning·Columbia University (Columbia+)

Supervised & unsupervised ML, linear algebra, neural networks, Transformers.

›

Prompt Engineering & Programming with OpenAI·Columbia University (Columbia+)

LLM prompt design, LangChain integration, controlled text generation.

›

Generative AI for Data Science·Microsoft

Generative AI concepts, LLM applications, responsible AI in data science workflows.

›

Back-End with Node.js & Express·IBM

Server-side development, REST API design, middleware architecture.

Vetriselvan R

Loading

Vetriselvan R