sastra
tcs
AI Vertical · Academic & Clinical Impact

Six Strategic AI Research Thrust Areas

AI & ML works on AI that can be explained, deployed and trusted. Its research spans medical image classification and segmentation, clinical decision support, AI-assisted teaching and assessment, foundation and agentic models, and low-cost edge deployment. Faculty, Ph.D. scholars and student interns work together on shared DGX computing, and every major result is packaged as an application that end users can test.

INFOGRAPHIC PIPELINE

AI & ML Medical Intelligence & XAI Processing Workflow

How clinical raw scans transform into verified, interpretable diagnostic recommendations with visual heatmaps.

01
Multimodal Scan Acquisition

X-Ray, Ultrasound, Brain CT, Thermal & MRI Scans

DICOM Standard
➔
02
Preprocessing & ROI Segmentation

U-Net ROI extraction, artifact removal & radiomic metrics

Radiomics + GLCM
➔
03
Deep Feature Extraction

MaxViT, Swin Transformer & ResU-Net with Attention

DGX Spark Accelerated
➔
04
Explainability (XAI) & Reports

Grad-CAM, Eigen-CAM & natural language diagnosis

Clinician-Ready

AI for Healthcare

Non-invasive screening, segmentation, and decision support from medical images
MedTech Hub

The healthcare group designs and evaluates classification and segmentation models for X-ray, thermal, CT, MRI, histopathology, and ultrasound scans. Each algorithm is accompanied by visual explanations (Grad-CAM, Eigen-CAM) and packaged into testable clinician interfaces.

Focus Disciplines
Prenatal cardiac screening Bone age assessment Thermal breast imaging Stroke CT classification Brain tumour MRI segmentation Osteosarcoma histopathology Neurogenetic disorders
Current Projects & Milestones
  • Fetal Cardiology Centre of Excellence AI decision support for congenital heart defect screening with Tiny Hearts Hospital
    Data Collection
  • BoneSight & BoneScale (CAMBAA) Pediatric hand X-ray bone age estimation via MaxViT, CBAM, and XGBoost
    Completed GUI
  • Thermal Breast Cancer Detection U-Net breast ROI segmentation with radiomic GLCM/LBP features and LIME
    Methods Complete
  • Brain Tumour Segmentation PSO-optimised Attention ResU-Net on BraTS 2019 (Dice 0.871, IoU 0.798)
    App Built
  • Brain Stroke CT Classifier Haemorrhagic vs ischaemic stroke differentiation with Grad-CAM heatmaps
    App Built
  • Osteosarcoma Classification CNN–Swin Transformer on Hugging Face; Graph Neural Network integration
    In Progress

AI for Education (SARADHI AI & SASTRAISM)

Interactive pedagogy, automated assessment, and edge classroom servers
SARADHI AI

Tools designed to maximize student engagement while saving teacher bandwidth. SARADHI AI's SARA platform runs live polls, multilingual quizzes, and AI-assisted grading. TARA products actively probe students during lectures, automate answer-key evaluation, and facilitate coding practice through CoderPlay AI.

Focus Disciplines
Live classroom interaction AI-assisted assessment Multilingual audio quizzes Coding education (11 langs) Offline school servers NEET/JEE preparation Teacher capacity building
Current Projects & Milestones
  • SARADHI AI Teaching–Learning Platform Cloud Kubernetes deployment; active student and faculty pilot trials
    Pilot Active
  • Audio Quiz Spark Multilingual interactive audio quiz generator, online and offline support
    Deployed
  • SARADHI AI SchoolBox Offline-first classroom server deployed on Jetson Orin at Sandipani School
    Field Deployed
  • AI Probing Engine Adaptive contextual questions for every student during a live classroom session
    Testing
  • AI Script Correction Automatic handwriting OCR and key-based evaluation of answer sheets
    In Progress
  • CoderPlay AI Coding platform supporting 11 languages with AI-verified certifications
    Deployed at SASTRA

Generative & Agentic AI

In-house foundation model KALI, domain reasoning LLMs, and RAG agents
Foundation LLM

Developing specialized reasoning models trained to power SARADHI AI and industrial verticals. KALI/SARAS (1.5B parameters) is under active pre-training. Specialized models for agriculture, strategic planning, chip design, disaster mitigation, and medical reasoning are undergoing domain fine-tuning.

Focus Disciplines
1.5B Parameter Foundation Models Domain-specific reasoning LLMs RAG architectures Autonomous classroom agents Generative CVAE design Speech interfaces
Current Projects & Milestones
  • KALI / SARAS Foundation Model 1.5-billion-parameter autoregressive language model under active DGX training
    Active Training
  • Specialized Domain LLMs AgriSenseLLM, StrategicLLM, ChipReasoningLLM, DisasterLLM, MediReasoningLLM
    R&D Phase
  • ChessGambitLM Long-horizon strategic reasoning and opening exploration engine
    R&D Phase
  • Satellite Origami Designer CVAE crease-pattern generator paired with RAG satellite design assistant
    Completed
  • Smart Spending Financial Assistant OCR receipt parser, transaction classification, and RAG budgeting agent
    Completed

Edge & Cognitive AI

Offline inference on Jetson hardware & 32-channel EEG cognitive workload estimation
Edge Hardware

AI that thrives without reliable internet connectivity, and AI that models the cognitive state of human operators. SARADHI AI SchoolBox embeds quizzes, teacher assistants, and learning analytics onto NVIDIA Jetson Orin devices, preserving local student data privacy. EEG models classify workload for pilot and operator safety.

Focus Disciplines
NVIDIA Jetson Orin NX Zero-cloud local inference 32-channel clinical EEG CNN–GCN hybrid models Aviation safety
Current Projects & Milestones
  • SARADHI AI SchoolBox Jetson Orin NX edge server deployed with local quiz generation in under 90s
    Deployed
  • CADAI Cognitive Workload Pilot mental fatigue and attention modeling from 32-channel EEG streams
    Completed
  • GenLife AI Edge + Generative AI assistant for neurogenetic and Parkinson's diagnosis
    In Progress

Agriculture, Aerospace, Oceans & Society

Applied physical AI: vernacular farm advisory, biomimetic robotics, and satellite design
Physical AI

Deploying AI solutions directly into field environments: voice advisory in the farmer's native regional language, autonomous biomimetic marine clean-up robots, and deployable structural space design. Systems are engineered for delivery via Data-as-a-Service and SaaS models.

Target Sectors
Agriculture Defence & Aviation Spacecraft Engineering Marine Clean-up Disaster Response
Current Projects & Milestones
  • AgriSahayak Voice-first advisory in English; Tamil, Hindi, Telugu, and Malayalam integration
    In Progress
  • Autonomous Ocean-Waste Whale Biomimetic marine drone collecting floating surface and seafloor micro-waste
    R&D Phase
  • CADAI Drone Cognitive workload and telemetry extended to autonomous search-and-rescue
    R&D Phase

Explainable AI & Decision Support

Interpretable heatmaps, feature attribution, and readable diagnostic narratives
XAI Suite

Every clinical and high-stakes model shipped by AI & ML exposes its underlying reasoning. Feature attribution visualizes pixel contributions in images, tabular models explain critical boundary rules, and automated clinical summaries ensure that doctors and patients can inspect predictions before acting.

Core Methods & Toolkits
Grad-CAM Eigen-CAM LIME SHAP Captum Attention Maps Range-Based Classifier
Current Projects & Milestones
  • CHD Risk Heatmap Reports Normal vs. Suspected vs. Critical congenital heart defect categorization
    Clinical Phase
  • Patient-Friendly Tumour Reports Grad-CAM localization mapped to clear, natural-language patient summaries
    Streamlit App Built
  • Range-Based Classifier Interpretable rule-based classifier competitive with Random Forest & KNN on medical data
    Validation
Copied to clipboard!