ANANTH
SHYAM

Who I am

Neural networks · Machine learning · AI systems

I work on neural networks, build machine learning models, and turn them into AI systems that hold up in the real world. Currently an M.Tech student in AI & Data Science at IIIT Bangalore.

• Neural network research at RISE Lab, IIT Madras
• Speech and language AI at AI4Bharat, IIT Madras
• Machine learning in industry: demand forecasting at XFormics Technologies
• AIR 235 in GATE DA 2026

• Neural networks
• Machine learning
• AI systems

What I work on

Neural networks · Machine learning · AI systems

Three things: neural networks that are accurate, robust and efficient; machine learning models that solve real problems; and AI systems, from RAG assistants to agentic workflows.

• Neural networks: knowledge distillation, quantization, adversarial robustness
• Machine learning: forecasting, stacked ensembles, anomaly detection
• AI systems: LLM and RAG assistants, agentic systems, speech alignment for 22 Indic languages

• Deep learning
• Model compression
• Robustness
• LLMs

Skills

PyTorch · TensorFlow · Keras

The languages, frameworks and tools I work with.

• ML & deep learning: PyTorch, TensorFlow, Keras, scikit-learn
• GenAI & agents: LLMs, RAG, LangChain, agentic systems, Hugging Face
• Languages & databases: Python, SQL, C, C++, R
• Data engineering: Spark, Kafka, Hadoop, MongoDB
• Tools: Azure, Docker, Git, Linux, Power BI, ONNX

• PyTorch
• TensorFlow
• Keras
• scikit-learn
• LangChain
• Hugging Face

M.Tech, Artificial Intelligence and Data Science

International Institute of Information Technology, Bangalore · July 2026 – June 2028

Currently pursuing an M.Tech in AI & Data Science.

• In progress

B.Tech, Artificial Intelligence and Data Science

Shiv Nadar University, Chennai · Sep 2022 – June 2026

Graduated with a CGPA of 9.019 / 10.

• CGPA 9.019 / 10

Research Intern

RISE Lab, IIT Madras · Chennai · June 2025 – Present

Developed QUARK, a unified framework for accurate, robust and efficient CNNs.

• Developed QUARK, a unified framework combining multi-teacher knowledge distillation, sensitivity-based mixed-precision quantization, and adversarial training to jointly optimize accuracy, robustness, and efficiency in CNNs
• Achieved up to 85.5% smaller model size, 37.15% lower energy use, and 46.79% higher PGD accuracy over baselines, outperforming SOTA methods on CIFAR-10/100 and ImageNet

• PyTorch
• Python
• Neural Networks
• CNN

RISE Lab

Data Science Intern

XFormics Technologies India · May 2025 – July 2025

Time-series demand forecasting on retail sales data.

• Built and benchmarked multiple time-series forecasting models (MQCNN, N-BEATS, GBM) on retail sales data, engineering SKU-level features to improve demand prediction accuracy
• Designed a stacked ensemble of MQCNN (deep learning) and XGBoost (machine learning), which gave the best demand forecasts, supporting inventory planning and business decisions

• Python
• ML
• DL
• Azure

XFormics

Research Intern

AI4Bharat, IIT Madras · Chennai · May 2024 – Sept 2024

Improved text-to-speech alignment for any language, applied to 22 Indic languages.

• Devised a novel methodology to improve alignment of Text-to-Speech (TTS) models for any language by extending a Montreal Forced Aligner (MFA) model, replacing the conventional g2p pipeline with a g2g approach for finer alignment
• Applied the methodology to 22 Indic languages, demonstrating improved accuracy over current legacy systems with minimal time difference
• Made open-source; received appreciation from Dr. Mitesh Khapra

• Python
• Linux
• NLP
• Transformers

AI4Bharat IndicMFA on GitHub

Lightweight and Efficient CNN

PyTorch · Knowledge distillation · Quantization-aware training · ONNX

Small, fast image classifiers for edge devices: an ensemble of large teacher models distilled into a tiny student, then quantized to INT8.

• Distilled an ensemble of ResNet18, MobileNetV2 and DenseNet121 teachers into a MobileNetV3-Small student, trained with quantization-aware training (cross-entropy + distillation loss under simulated INT8)
• Rice varieties (5 classes): a 0.31 MB INT8 model at 98.07% accuracy, 3.95 ms per image on CPU
• Rice diseases: 3× smaller and 3× faster after INT8 quantization (3.32 MB → 1.10 MB, 178.8 → 529.5 FPS)
• Exported to ONNX and demoed in a mobile app (video on GitHub)

• Knowledge distillation
• Quantization
• Edge AI
• Computer vision

GitHub

Lightweight and Efficient YOLOv5

PyTorch · YOLOv5 Nano · Post-training quantization · ONNX

A compact object detector for rice quality inspection on edge devices: YOLOv5 Nano trained on 8 classes, then quantized to INT8 with no loss in mAP@50.

• Trained YOLOv5 Nano to detect 8 classes of rice grains and contaminants (broken, chalky, unsound, stone, plastic and more) from the Roboflow Rice Quality dataset
• Static post-training quantization from FP32 to INT8 ONNX: 2.7× smaller (8.06 MB → 2.99 MB)
• mAP@50 held at 0.885 after quantization, with precision up from 0.834 to 0.855
• Exported to ONNX and demoed in a mobile app (video on GitHub)

• Object detection
• Quantization
• Edge AI
• Computer vision

GitHub

Araycci, the research paper AI assistant

Python · LLAMA3 · Hybrid RAG · Hugging Face

Lets users interact with 2.4 million research papers from arXiv.

• Uses LLAMA3 and a hybrid RAG pipeline (BM25 retrieval with cross-encoder reranking) for clustering, summarization and multilingual Q&A
• Coordinated the team and designed the retrieval architecture, improving search relevance over pure dense retrieval, especially for keyword- and entity-heavy queries

• RAG
• LLMs
• NLP

Live demo GitHub

Advanced Momentum Stock Model

Python · yfinance · Technical indicators · Backtesting

A stock screener that scores NSE-listed stocks on momentum signals. In backtests over the past 5 years it beat the NIFTY 50 70% of the time.

• Scores each stock against nine conditions built from RSI, MACD, ADX, the stochastic oscillator, relative volume and recent price trend
• Backtested on yfinance data over the past 5 years, outperforming the NIFTY 50 benchmark 70% of the time
• Companion tools to explore one stock in depth (valuation, profitability, growth, analyst recommendations, price charts) or compare several side by side

• Quantitative finance
• Technical analysis
• Backtesting

GitHub

CarbonAware-BMS

Python · Azure ML · Spark · Kafka · Power BI

A carbon-aware building management system that streams data from 100 building sensors, forecasts the next 24 hours of energy use, flags anomalies and suggests greener times to run loads.

• Streaming pipeline from 100 IoT sensors (aggregated every 5 minutes) through Azure IoT Hub, Azure Functions, Azure SQL and Data Lake, with Kafka and Spark for streaming and batch processing
• 24-hour energy forecasting with XGBoost and MQCNN, including 95% confidence intervals
• Isolation Forest anomaly detection to catch energy waste and equipment degradation early
• Live carbon tracking (kWh → kgCO₂e) with load-shifting recommendations in Power BI, and automatic model retraining on Azure ML

• Azure
• Forecasting
• Anomaly detection
• IoT

GitHub

GATE DA 2026 · AIR 235

2026

All India Rank 235 in GATE Data Science and Artificial Intelligence.

• AIR 235

Finalist, SSN HackIT 2023

Inter-college hackathon · powered by Google

Reached the finals of the SSN HackIT inter-college hackathon.

• Hackathon

NTSE Finalist and Olympiad medals

School

NTSE finalist, with multiple state-level medals in the SOF, IAPT and PRMO Olympiads.

• NTSE
• Olympiads

University football team

Shiv Nadar University, Chennai

Represented SNU Chennai in the South India Zonals and other tournaments.

• Football

SIP State Level Abacus Competition

Tamil Nadu

Placed 3rd at the Tamil Nadu state level, twice.

• 3rd place × 2