Experience

Founding Engineer — LLM Evaluation Infrastructure

SUAV Tech · 2025 – Present

• Designed end-to-end LLM evaluation pipelines across 15+ risk categories, reducing undetected violations 40% pre-deployment.
• Built modular Python harness enabling new benchmark integration in under 2 hours.

WebOps Secretary

Student Executive Council, IIT Madras BS · Jan 2026 – Present

• Built and shipped a production RAG chatbot serving 40,000+ students with a daily CI/CD sync pipeline using hash-based incremental chunk invalidation.
• Engineered a multi-layer query pipeline — NER-keyed Redis cache, semantic cache, and Qdrant hybrid search (dense + sparse + RRF) — optimizing for latency and cost while maintaining recall.

Power Load Forecasting System

LightGBM · FastAPI · Docker · GCP · Next.js

Time-series load classification on 35K+ records at 430 RPS — LightGBM with cyclical feature engineering and GCP deployment.

GitHub · Live Demo

Genrefy — Audio Classifier

PyTorch Lightning · EfficientNet · W&B · Streamlit

EfficientNet-V2 trained on log-mel spectrograms for 10-genre music classification under distribution shift (clean stems → noisy mashups) — 0.930 on Kaggle private LB.

GitHub · Live Demo

DeepWaste — CNN vs Fine-tuning Benchmark

PyTorch Lightning · ONNX · TensorRT · W&B

Deployment benchmark across 5 inference formats — TensorRT FP16 achieved 6,905 img/s with 98.41% accuracy on garbage classification.

GitHub

Sales Forecasting — Kaggle Rank 133

Python · Scikit-learn · Fourier Analysis · Streamlit

Multiplicative forecasting model with Fourier analysis and GDP indicators — 12.7% MAPE across 6 countries, 10 years of data.

GitHub · Live Demo

Customer Profiling

Python · Machine Learning · Clustering · K-Means · DBSCAN

K-Means & DBSCAN clustering algorithms for customer segmentation analysis with interactive profiling dashboard.

GitHub · Live Demo

CrossEmotion — Multimodal Emotion Recognition

PyTorch · R3D-18 · BERT · Conformer · AWS SageMaker

Fusion of video (R3D-18), text (BERT), and audio (Conformer) with cross-modal attention — 67.8% accuracy on MELD dataset.

GitHub

Logic-Guided Diagnosis — Neuro-Symbolic AI

PyTorch · Decision Trees · Scikit-learn

Residual MLP trained with differentiable rule-agreement loss — 88.32% accuracy, 0.9516 ROC-AUC on heart disease prediction.

GitHub