↵ enter

MS Applied Statistics · NYU · New York

Aravind
Kannappan

ML researcher and engineer at the intersection of AI safety, Bayesian statistics, and applied deep learning. Currently a research intern at NYU Grossman School of Medicine and a fellow at the Supervised Program for Alignment Research.

Previously founded Synthure (clinical LLM) and currently building Replays AI (AI sports commentary). Open to relocation.

AI Safety Bayesian Methods NLP Computer Vision Genomics Reinforcement Learning PyTorch CUDA

Technical Blog

Experience

  • ML Research Intern
    NYU Grossman School of Medicine · Feb 2026 – Present · New York, NY
    Trained VAE and transformer models in PyTorch to predict functional variant effects across 100+ genomic tracks. Developed hierarchical Bayesian VI integrating ATAC-seq and ChIP-seq signals. AUROC 0.83 on eQTL prediction, +11% over LASSO.
  • AI Safety Research Fellow
    Supervised Program for Alignment Research · Feb 2026 – Present · Remote
    Designed evidence-weighted evaluation frameworks. Implemented simulation-based inference with Set Transformers over 500K trajectories. Applied mechanistic interpretability across 70M–12B parameter models to identify confidence divergence and error modes.
  • Software Engineer AI/ML · Technical Lead
    Replays AI · Sep 2025 – Present · New York, NY
    Built LLM-based sports commentary system (+45% engagement). Designed RAG pipeline over 5M+ events with hybrid sparse/dense retrieval. Reduced inference latency 850ms → 340ms via INT8 quantization and KV-cache reuse at 10K+ scale.
  • Founder & CTO
    Synthure · Jun 2023 – Jul 2025
    Clinical LLM over 60K patient records. Reduced hallucination rate 81% → 13% via SFT + RLHF. Domain-specific entity recognition cut policy misinterpretation by 94%. Measured bias across protected classes to ±5% variance.

Projects

  • GameNGen
    Generative Models · Game AI
    Real-time neural game engine. Replicates DOOM gameplay frame-by-frame using a diffusion model trained on game trajectories — no explicit game logic, pure learned simulation.
    PyTorchDiffusionRLCUDA
  • Volatility Alchemist
    Quantitative Finance · Time Series
    Options volatility surface modeling with GARCH, stochastic vol models, and ML-based regime detection. Calibrates implied vol surfaces to market data and forecasts realized volatility across regimes.
    PythonGARCHTime SeriesXGBoost
  • Traffic Flow Optimization
    Reinforcement Learning · Graph Theory
    RL agent for adaptive traffic signal control modeled as an MDP over a city graph. PPO-trained, outperforming fixed-time baselines by 23% in average wait reduction across simulated intersections.
    PPOGNNPyTorchSimulation
  • RobustSight
    AI Safety · Computer Vision
    Evaluation framework for CV model safety and alignment. Benchmarks adversarial robustness (PGD, FGSM, patch attacks), saliency interpretability, and human alignment on CIFAR-10/C.
    PyTorchAdversarial MLInterpretabilityCIFAR-10