Advancing AI & Computer Vision Architecture.
Specializing in Few-Shot Domain Adaptation, 3D Cryo-ET Subtomogram Classification, and building robust full-stack deep learning engines.
+3-Shot
Classification Gain
Swin3D
Backbone Adaptation
CMU / NTU
Global Institutions
Full-Stack
AI Applications
Research & Experience
Pioneering research across premier global institutions in computer vision, subtomogram classification, LLM alignment, and deep learning.
Carnegie Mellon University
Researcher @ Xu Lab
- Developed a Hierarchical Domain Adaptation framework for few-shot Cryo-ET subtomogram classification, integrating Spatial Transformer Networks (STN), Difference-of-Gaussian intensity filters, and color transforms with MMD and CORAL alignment losses on Swin3D backbones.
- Engineered extensive parameter sweeps across 10 random seeds and low-shot target datasets (Noble & Qiang), outperforming target-only baselines and boosting 3-shot accuracy by up to +6.05%.
- Optimized few-shot generalization in 5-shot regimes, achieving 63.50% accuracy on Noble and 58.83% on Qiang (+7.16% over target-only models) while preventing optimization collapse under strict low-data constraints.
Nanyang Technological University (NTU)
Research Intern @ VIDA Lab
- Built and released a large-scale infographic generation dataset by scraping open-source repositories, performing structural curation, and parsing semantic image-instruction pairs for diffusion model pretraining (published on Kaggle and cited in independent research InfoAffect, 2025).
- Developed and optimized an LLM-driven instruction synthesis pipeline leveraging prompt engineering to produce hierarchical, step-wise infographic construction instructions.
- Synthesized 50 papers (IEEE VIS, IEEE TVCG, ACM CHI) to frame the methodological landscape of automated infographic generation for Stable Diffusion fine-tuning strategies.
NITK Surathkal
Computer Vision Research Intern @ HALE Lab
- Fine-tuned and aligned Ornith 9B and 35B parameter LLMs, improving downstream task accuracy by 24% and reducing preference alignment compute costs by 50% using multi-stage SFT and Direct Preference Optimization (DPO).
- Scaled test-time reasoning capacity across long-horizon outputs (>20k tokens), increasing complex problem-solving accuracy by 18% using Group Relative Policy Optimization (GRPO) and dynamic entropy-regularized chunk masking.
- Optimized high-throughput model inference across multi-GPU setups, reducing peak VRAM consumption by 40% using custom 2D tensor auto-scaling and dynamic context pruning.
Mohamed bin Zayed University of AI (MBZUAI)
Summer Research Intern
- Designed and deployed a semi-supervised deep clustering framework for 3D Cryo-ET tomograms, integrating a YOPO-based feature extractor with a vectorized GMM initialized via k-means and label constraints.
- Engineered a PyTorch pipeline supporting modular feature extraction, latent space modeling, and dynamic label refinement, improving clustering fidelity on low-SNR volumetric datasets.
- Evaluated learned representations across multiple cryo-ET benchmarks, demonstrating improved compactness and class separability in feature space for subtomogram anomaly detection.
Carnegie Mellon University
Research Assistant @ Xu Lab
- Under Dr. Min Xu, built complete processing pipelines for Cryo-EM and Cryo-ET volume reconstruction utilizing EMAN2, CryoSPARC, and Warp for subtomogram averaging and 3D classification.
- Automated high-throughput analysis across hundreds of tomograms, ensuring reproducibility in ultrastructural reconstruction of macromolecular assemblies.
IRIS, NITK
Machine Learning Developer
- Deployed production ML modules into IRIS—the university ERP platform serving 7,000+ staff and students.
- Developed a profanity detection and sentiment analysis system leveraging pretrained transformer-based models and rule-based augmentations.
- Used Ollama LLaMA 7B model to build a summarization tool for 10+ years of placement data, predicting key skills linked to successful hiring outcomes.
Education
Bachelor of Technology (B.Tech) - Information Technology
2023 - 2027National Institute of Technology Karnataka (NITK), Surathkal
Focus on Computer Vision, Machine Learning, Natural Language Processing, Deep Learning, and Reinforcement Learning.
Featured Projects & Systems
A comprehensive showcase of deep learning architectures, computer vision pipelines, open-source repositories, and full-stack web applications.
Hierarchical Domain Adaptation for Cryo-ET
Few-shot subtomogram classification framework integrating STN, DoG filters, MMD, and CORAL alignment on a Swin3D backbone to bridge synthetic-to-real domain gaps.
Twinscribe / Infographics LLM & Diffusion Pipeline
Large-scale infographic generation dataset & LLM-driven instruction synthesis pipeline for Stable Diffusion text-to-infographic supervision.
Vision Kinect (Gesture Tetris)
Hands-free Tetris application utilizing YOLOv5 object detection and OpenCV real-time hand-gesture recognition wrapped in a web interface.
Imagined Risk (World Models & Autonomous Driving)
Geometric world models and multi-view vision-language reinforcement learning for autonomous driving perception and risk-imagination reasoning.
3D CNN Video Classification Engine
3D Convolutional Neural Network architecture in PyTorch for spatial-temporal action recognition and video clip classification.
Interactive 3D ML Computation Graph Portfolio
High-performance interactive 3D WebGL computation graph visualizing foundation model interpretability, attention routing, and diffusion noise.
Professional Full-Stack Developer Platform
Custom developer platform featuring embedded interactive shell, API telemetry endpoints, and dark-theme glassmorphism UI.
Irish Poem Generator
LSTM neural network predicting poetic text from seed phrases utilizing embedding layers and dense word representations.
Lunar Lander Reinforcement Learning
Deep Q-Learning agent trained to achieve soft landing of a lunar module on designated pads while minimizing impact velocity.
Research & Academic Contributions
Detailed focus on domain adaptation frameworks, 3D bio-imaging classification, and deep neural network evaluation.
Hierarchical Domain Adaptation for Few-Shot Cryo-ET Subtomogram Classification
Bridging the synthetic-to-real domain gap inCryo-Electron Tomography subtomograms by jointly optimizing input-level spatial & color transformations with feature-level MMD & CORAL losses on Kinetics-400 pretrained Swin3D backbones.
Multimodal Representation Interpretability in Deep Vision Models
Analyzing feature map activations and attention routing matrices in vision-language models to evaluate transferability across low-shot downstream benchmarks.
Skills & Technical Capabilities
A comprehensive breakdown of expertise across AI research, computer vision frameworks, and full-stack web architectures.
AI & Machine Learning
Computer Vision & Deep Learning
Full-Stack & Web Engineering
Backend, DevOps & Systems
Let's Build Next-Gen AI Together
Open to research collaborations, AI/ML engineering roles, computer vision consultations, and full-stack projects. Feel free to send a direct transmission or connect via professional networks.
Send Transmission
CONNECTED TO NEXT.JS FULL-STACK API ROUTE