AI Researcher • Engineer • Patent Holder
AI should not just predict. It should understand.
Bridging perception, reasoning, and emotional intelligence
in intelligent systems.
About
I am an AI researcher and engineer focused on building intelligent systems that are accurate, interpretable, and deployable. My work spans computer vision, natural language processing, explainable AI, and agentic systems.
Indian Patent holder. German Research Grant recipient. Published at ASCE 2025 in Arizona. UCL London Summer School alumna. Exchange scholar at ESIGELEC, France. Interned at Deloitte India. I bridge research and production, transforming ideas into working AI systems.
Education
B.Tech Computer Science — Manipal University Jaipur
Expected 2027 • CGPA 9.84/10 • Dean's List
Global Exposure
UCL London • ESIGELEC France
Data-Driven Web Apps & AI • Automotive Embedded Systems
Focus Areas
CV • NLP • XAI • GenAI • Agentic AI • MLOps
9.84
CGPA / 10
4x
Dean's List
2
Patents / Grants
4
Research Papers
Recent Interests · Articles
Short, opinionated pieces on the frontier I'm obsessed with — ethics, interpretability, healthcare XAI, and the new ML edge in quant. Each one ships with an interactive visualization.
Ethics Compass
Game-theoretic governance vs. laissez-faire vs. static guardrails.
Anthropic's Claude refused to ghostwrite a viral fanfic. Mythos got jailbroken into impersonating a real teen. Fable shipped a celebrity deepfake before lunch. The same week. Same news cycle. Same wake-up call: alignment is not a feature, it's the product.
Read & interactSparse Features
SAE-decoded monosemantic clusters: syntax · entities · math · code · deception.
LLMs and RAGs both stand on one backbone: the transformer. We treat it as a black box because the residual stream is dense and polysemantic. Anthropic's sparse autoencoder work showed those tangled features actually live on a clean geometric manifold — surprisingly close to how humans organize meaning.
Read & interactXAI Backtrace
Prediction → layer → attention → Grad-CAM → clinician trust.
94% accuracy doesn't move a single clinician if they can't see why. XAI is the chain of custody between a model prediction and a radiologist's pen. Grad-CAM, LIME, SHAP and integrated gradients aren't cosmetics — they're the audit trail medicine has demanded since the Hippocratic oath.
Read & interactFactor crowding has crushed classical alpha. Transformer-based regime detection, diffusion priors on order flow, and offline RL execution policies are the new edge — each compounding while the old edge melts.
Classical factor alpha has been crowded into oblivion. The 2026 edge is built from offline RL execution, transformer-based regime detection, diffusion priors over order-flow, and SAE-interpretable feature stores. The quants who treat ML as a black box are about to get arbitraged by the ones who treat it as a microscope.
Read & interactExperience
Technology & Transformation Intern
Jun 2025 — Jul 2025
Delhi, India
Technical Apprentice
May 2021 — Jun 2021
Delhi, India
Research & Patents
ASCE International Conference on Transportation & Development — Glendale, Arizona, USA • Jun 2025
View all on ResearchGateLead modeling and simulation. Implemented RBF networks for time-series price diffusion.
Co-designed fuzzy inference system. Applied K-means clustering for budget forecasting.
Built regression framework with bootstrapped sampling for variance-aware predictions.
Survival analysis-based hazard model for time-to-milestone risk prediction.
Projects
Tap any project to dive into architecture, results, images, demo video and the GitHub repo.
Computer Vision Portfolio
From medical imaging to real-time detection — building vision systems that are accurate, interpretable, and deployment-ready.
Featured Research
Medical Imaging / XAI
Alzheimer's & Parkinson's Classification from Brain MRIs
Overview
A lightweight multi-task deep learning framework that simultaneously classifies Alzheimer's disease stages and Parkinson's disease from brain MRI scans, integrated with Explainable AI for clinical trust and transparency.
Problem
Current diagnostic pipelines treat Alzheimer's and Parkinson's as separate classification tasks, requiring multiple models, redundant computation, and inconsistent interpretability — limiting real-world clinical deployment.
Technical Approach
Architecture Pipeline
Key Results
94.2%
Alzheimer's Accuracy
96.8%
Parkinson's Accuracy
~40%
Model Size Reduction
<200ms
Inference Time
Techniques & Tools
Impact
Indian Patent filed (Jul 2025). German Research Grant awarded (Oct 2025). Framework enables unified, interpretable neurodegenerative diagnosis — bridging the gap between deep learning accuracy and clinical trust.
Applied Vision Projects
Hands-on implementations spanning gesture recognition, real-time tracking, augmented reality, and intelligent monitoring systems.
Real-time red color detection using webcam input and HSV color space segmentation. Dual HSV thresholds handle hue wrapping with morphological filtering for noise reduction. Displays original frame, mask, and segmented result simultaneously.
Extended red color detection with image moments to compute the centroid of detected regions in real time. Tracks object motion via dynamic centroid overlays, contour filtering, and moment-based coordinate computation for stable tracking.
Interactive gesture-based balloon popping game using real-time hand tracking. MediaPipe detects fingertip landmarks mapped to a virtual game canvas with collision detection, animated sprites, scoring logic, and leaderboard storage.
Real-time gesture-based drawing interface using centroid tracking of a colored object. Computes centroids via image moments, renders motion trails on a virtual canvas with color selection zones and interactive painting controls.
Adaptive background modeling system using running average for motion detection. Gaussian blur and morphological filtering minimize noise. Triggers asynchronous alarms and auto-saves timestamped snapshots during intrusion events.
Augmented reality illusion detecting a specific cloak color via HSV segmentation. Generates masks to isolate cloak regions and replaces them with a previously captured background frame with morphological smoothing for seamless blending.
Touchless virtual piano controlled entirely by fingertip gestures. MediaPipe hand tracking detects finger positions mapped to key regions as sound triggers with cooldown constraints and threaded audio playback.
Touchless dataset explorer using swipe and pinch gestures. Wrist displacement tracking for navigation and Euclidean distance measurement for zoom control with dynamic image resizing and highlighted thumbnails.
Vision-based virtual mouse where fingertip movements control the cursor. Pinch detection for click/drag operations with smoothing filters to reduce jitter, mapped from webcam coordinates to full-screen resolution.
YOLO-based object detection for static images and live webcam streams. YOLOv8 processes frames for bounding box prediction and classification with labeled real-time outputs for surveillance, automation, and smart monitoring.
Real-time environmental monitoring combining YOLO object detection with lifecycle emission modeling. Assigns ecological weights using DEFRA/EPA/IPCC datasets, computes per-object impact with logarithmic normalization, and aggregates into a scene-level Eco Index (0–100).
“Vision is not just seeing — it's understanding what you see.”
Skills & Recognition
Languages
AI / ML
Web & Backend
Tools & Infra
Recognition & Awards
Contact
Open to AI/ML research collaborations, product-driven engineering roles, and opportunities that demand ownership and impact.
SP.
Designed with precision. Built to last.