F-01
Artificial Intelligence · Foundations
Foundational Intelligence — Algorithms that Turn Complexity into Insight
Objective
Advance general-purpose learning and representation algorithms that turn high-dimensional, messy real-world signals into reliable, actionable insight — the shared methodological backbone for every other frontier in the lab.
Hypothesis
Progress depends not on scale alone but on architectures that compose perception, memory and reasoning under explicit objectives — making AI a practical enabler for hard, real-world problems.
Approach
Cross-domain deep learning, multimodal and representation learning, hybrid feature engineering, and careful, reproducible benchmarking across medical, industrial and scientific data.
Deep LearningRepresentationMultimodalBenchmarks
F-02
Quantum × Medical AI
Q-HyFuse — Quantum-Annealed Fusion of Deep & Radiomic Biomarkers
Objective
Select a compact, low-redundancy biomarker subset for lung-cancer malignancy from a fused feature space of CNN (InceptionV3, 2048-d), ViT-B/16 (768-d) and PyRadiomics (~107-d), using quantum-annealing QUBO mRMR.
Hypothesis
Quantum annealing can recover more compact, less-redundant biomarker subsets than greedy classical mRMR — without losing diagnostic accuracy — because the QUBO encodes relevance and redundancy jointly.
Approach
QUBO mRMR (relevance − redundancy + cardinality) solved across paradigms: simulated annealing & exact solver, gate-model QAOA (Qiskit Aer), with hooks for D-Wave Advantage and IBM hardware. Scoped honestly — no claim of quantum advantage.
QUBOQAOAD-WaveRadiomicsSHAP
F-03
Computational Imaging · Inverse Problems
Robust QPAT via Correction-Augmented, Matrix-Free Gauss–Newton
Objective
Make quantitative photoacoustic tomography (QPAT) reconstruction stable and uncertainty-aware under optical model error — without ever storing the multi-gigabyte Hessian.
Hypothesis
A learned, depth-aware model-error correction operator paired with a matrix-free preconditioned Gauss–Newton iteration yields calibrated reconstructions whose predicted uncertainty correlates with true error.
Approach
Exact matrix-free Jacobian (J·v via one sparse solve), learned ridge fluence correction, deep-ensemble heteroscedastic uncertainty, validated discretisation-agnostically on FD and unstructured P1 Delaunay FEM meshes.
Gauss–NewtonFEMUncertaintyInverse problem
F-04
Medical Image Analysis
Explainable Deep Assessment of Lung-Nodule Malignancy
Objective
Build high-performance, clinically interpretable CNN systems that estimate lung-cancer risk from thoracic CT — advancing accurate, robust and explainable disease detection. The core of the PhD thesis at NIT Raipur.
Hypothesis
Hybrid CNN feature engineering with explainable-AI overlays improves clinician trust and decision support without trading away classification accuracy.
Approach
Systematic hyperparameter optimisation, ConvMixer architectures, hybrid feature fusion and XAI attribution — work that has appeared in Physics in Medicine & Biology and Physica Medica, and received an IEEE best-paper presentation award (SPA 2023, Poznań).
CNNConvMixerExplainable AIThoracic CT
F-05
Biomedical Optics · Simulation
A Physically-Grounded Photoacoustic Tomography Digital Twin
Objective
Build a full forward + inverse PAT pipeline — Monte-Carlo / diffusion optics, FEM acoustics, and back-projection reconstruction — as a reproducible benchmark for imaging priors.
Hypothesis
A physically faithful digital twin (chromophore-based phantoms, Henyey–Greenstein scattering, Newmark-β acoustics) lets reconstruction priors be compared fairly, independent of real-hardware noise.
Approach
A modular Python lab: HbO₂/Hb spectra → μₐ(λ), quad/tri/Delaunay meshes, explicit & implicit solvers, filtered universal back-projection, and wave-propagation movie export. Grounded in Wang & Hu (Science 2012), MCML and k-Wave.
Monte CarloFEM acousticsBack-projectionDigital twin
F-06
Robotics · UAV & Edge AI
Real-Time Edge Perception for Robotics & Autonomous UAVs
Objective
Run accurate detection & perception models on sub-10-watt embedded hardware (NVIDIA Jetson Nano) for real-time robotic and aerial inspection — work conducted as Principal Project Associate at IIT Bhilai.
Hypothesis
Model compression and sensor fusion can preserve detection quality while meeting the latency and power envelope of airborne and mobile-robot edge devices.
Approach
YOLO-class detectors, ONNX/TensorRT optimisation, LiDAR + camera sensor fusion, and on-device real-time inference — the perception backbone behind NAIQ's industrial inspection work.
Jetson NanoYOLOTensorRTSensor fusion
F-07
Autonomous Mobility · Perception
Perception & Decision Stacks for Autonomous Vehicles
Objective
Build robust real-time perception, sensor-fusion and decision pipelines for self-driving and autonomous ground vehicles operating in unstructured, safety-critical environments.
Hypothesis
Tightly-fused multi-sensor perception with uncertainty-aware decision-making generalises to unstructured roads far better than camera-only or hand-coded rule-based pipelines.
Approach
Camera + LiDAR + radar sensor fusion, 3D object detection & multi-object tracking, drivable-space segmentation and real-time motion planning — sharing the edge-inference backbone of the robotics/UAV frontier.
LiDAR3D DetectionTrackingPlanning
F-08
Simulation · Real-Time Visualization
Simulated & Real-Time Scientific Visualization
Objective
Turn complex physical simulations and live sensor streams into interactive, physically-faithful 3D visualizations — digital twins that scientists and operators can explore and steer in real time.
Hypothesis
Coupling physically-grounded simulation with real-time, GPU-accelerated visualization closes the loop between model, measurement and human insight faster than static, offline analysis.
Approach
Real-time rendering (WebGL / Three.js, GPU shaders), physics-based simulation (Monte-Carlo optics, FEM acoustics, wave propagation) and interactive digital twins — for example the PAT wave-propagation visualizer and live edge-inference dashboards.
Real-time 3DWebGLDigital twinSimulation
F-09
LLM Alignment · RLHF
Human-Aligned Reward Modelling for Frontier LLMs
Objective
Engineer reinforcement-learning-from-human-feedback (RLHF) pipelines and evaluation harnesses that make large language models more truthful, helpful and instruction-faithful.
Hypothesis
Structured human-preference signals and calibrated reward models reduce hallucination and improve instruction-following more reliably than scale alone.
Approach
Preference-data design, reward modelling, model evaluation and red-teaming — experience built as an AI/LLM training specialist across enterprise LLM pipelines, informing NAIQ's own assistant layer.
RLHFReward modelEvaluationRed-team
F-10
Affective Computing · HealthTech
EI-OS — An Emotional-Intelligence Operating System for Wellbeing
Objective
Deliver scalable, privacy-preserving emotional-wellness support — starting with Indian students — through AI-driven affective assessment under NAIQ Health Tech.
Hypothesis
Multimodal affective signals, processed locally, can provide meaningful and trustworthy emotional support at scale without compromising privacy.
Approach
Local-first inference, multimodal signal modelling and longitudinal wellbeing tracking — designed around consent, data minimisation and on-device processing.
Affective AILocal-firstWellbeingPrivacy
F-11
AI for Society · Sustainability
AI, Humanity & Sustainability — Intelligence Aligned with Human & Planetary Value
Objective
Ensure advances in intelligence are matched by advances in ethics, sustainability and human values — measuring and reducing the social and ecological footprint of AI while supporting collective wellbeing.
Hypothesis
Technology advances best when innovation is balanced with ethics and sustainability — systems designed for energy efficiency, fairness and human dignity deliver more durable, real-world value than raw capability alone.
Approach
Energy-aware, edge-first compute; responsible-AI evaluation (bias, transparency, consent); and sustainability-oriented design principles embedded across every NAIQ product and frontier.
Responsible AISustainabilityEthicsEfficiency
F-12
Foundations of Intelligence · Theory
Self-Evolving Mathematical Frameworks for Consciousness
Objective
Explore unified mathematical structures for consciousness that move beyond knowledge-based, feed-forward computation — a long-horizon, exploratory frontier of the lab.
Hypothesis
Consciousness may be describable by a self-referential, self-evolving mathematical operator acting from deeper, universal principles — qualitatively distinct from the stateless function-approximation of today's neural networks.
Approach
Formal modelling of self-reference and recursion, dynamical-systems formulations, and bridges between information theory, physics and cognition. Openly speculative; pursued as foundational inquiry, not product.
TheoryDynamical systemsInformationCognition
F-13
Representation Geometry · Foundations
Semantic Compression Dynamics (SCD) — Intelligence through Compression, Geometry & Continuous Representation
Objective
Reformulate intelligence as the discovery, manipulation and communication of highly compressed semantic structures that evolve within continuous geometric spaces — treating compression, geometry and continuous representation as the substrate of understanding.
Hypothesis
Intelligence can be expressed as compression dynamics — learning, reasoning and communication reduce to finding and moving through low-dimensional semantic manifolds, where geometry, not symbolic rules, governs meaning.
Approach
Continuous latent representations and information-theoretic compression objectives (rate–distortion, minimum description length), manifold and geometric learning, and dynamical flows over semantic embedding spaces — bridging representation learning, information theory and differential geometry.
CompressionGeometryContinuous RepresentationInformation theory