Chirag Jain

Apollo Tyres Ltd. Digital Innovation Hub

AI & Data Science Intern · Technical In-Charge

January 2026 – Present

Location: Hyderabad

apollotyres.com ↗

Technical In-Charge on multiple AI initiatives aimed at operational efficiency in manufacturing.

Stack: Strands Agent, GraphRAG, Redis, XGBoost, LightGBM, CatBoost, Ridge, Classical CV, Python, LLMOps

• Production multi-agent AI system — Designed and deployed a production-grade multi-agent orchestration system using Strands Agent and GraphRAG (five specialized agents), live at the Limda Manufacturing Plant. Operators retrieve performance data and generate downloadable charts via natural language. Cut manual dashboard navigation effort by 40–50% with Redis caching, Chain-of-Thought tracing, human-in-the-loop validation, and robust fallbacks. Under iterative deployment from plant-side feedback.
• Energy demand forecasting & power bidding — Leading Phase 2 of an ensemble forecasting system on a production time-series model with dynamic recalibration (XGBoost, LightGBM, CatBoost + Ridge meta-learner). Phase 1 delivered ₹1 Cr+ cost savings in three months. Achieved median SMAPE 9.08% and average net bias +1.98%; used for real-time power bidding on the Indian Energy Exchange.
• Computer vision for quality inspection — Classical CV pipeline to remove background streaks from SEM images of rubber compounds. After evaluating 20+ methodologies, deployed directional morphological opening that cut impurity overestimation by 1–2 percentage points across 36 images from four compound sets. Inference 4–5 seconds per set (target < 10s). Validated with senior researchers at Apollo R&D, Enschede (Netherlands) on new compounds and SKUs.

Indian National Science Academy (INSA) Summer Research Fellow

May 2025 – Jul 2025

insaindia.res.in ↗

Stack: PyTorch, Computer Vision

Supervisor: Dr. Ramesh Venkadamchalam (Dept. of Mathematics, Central University of Tamil Nadu)

• Research on AI-based disaster damage assessment using the xView2 dataset (20,000 pre/post-disaster image pairs), focusing on building-level classification for HADR efforts.
• Built a scalable preprocessing pipeline (patch extraction, augmentation, majority-class undersampling) with a 94.4% dataset expansion (304,370 → 591,583 patches) and improved class balance.
• Developed a lightweight residual CNN (3.7M params, 28.71 MB) at 83.3% accuracy — matching ResNet18 (83.5%, 7M params) with 54.66% smaller model size for field deployment.

Comparative Analysis of DL Models for Brain MRI Tumour Detection

IEEE Publication

Stack: PyTorch, Computer Vision

Supervisor: Dr. Deepasikha Mishra (SCOPE, VIT-AP University)

• Comparative study of VGG16, VGG19, Xception, Simple CNN, and EfficientNet-Attn under standardized training conditions.
• Processed ~12,000 brain MRI scans from the MRI ND-5 dataset (IEEE Dataport) through transformation pipelines with comparative performance visualizations.
• EfficientNet hit 99.82% accuracy (external) and 97.45% (internal), validated with Nemenyi and Cohen's d tests.
• Presented at an IEEE international conference; published in the Scopus-indexed IEEE Digital Library.