Work on client-facing computer vision projects covering image segmentation, understanding, and image classification with production latency constraints.
Automate MLOps workflows for evaluation and retraining.
Improve model accuracy on large image datasets without breaking strict latency budgets.
Built TARS, a fully offline multimodal AI agent unifying speech (Vosk, WebRTC, Silero VAD), vision (YOLOv8, BLIP, Haar cascades), and language (Phi-4-mini-instruct) in a local pipeline with no cloud calls or external APIs.
Implemented LangGraph orchestration with dynamic node routing, custom concurrency management, and fault-tolerant health checks with auto-restarts.
Optimized large models for CPU-only inference on 16 GB RAM using quantized LLMs, achieving real-time performance without GPU acceleration.
Built GenexLearn, a Next.js e-learning PWA with course management, admin dashboards, authentication, dark mode, and an OpenRouter AI chat feature with per-user usage quota tracking.
Built an offline-first desktop app that shows global rocket launches on an interactive 3D globe, using live data from the Launch Library 2 API.
Designed a data sync pipeline that works within the API's strict rate limits, and bundled a seed database so the app has full launch data on first install, with no internet needed.
Set up automated testing and CI with GitHub Actions, plus a one-command Windows release pipeline.
Designed and deployed the AWS architecture for a cloud-native pipeline that performs real-time distributed Lucky Imaging, where telescopes stream short-exposure frames to the cloud for automated quality selection and stacking.
Used continuous streaming and edge-based pre-filtering so only high-quality frames are transmitted, reducing bandwidth and latency, and let multiple telescopes share one processing environment for collaborative observations.
Built a reverse proxy pipeline that classifies HTTP attacks using a Random Forest model.
Applied Bayesian hyperparameter optimization and SMOTE for class imbalance, and added SHAP-based feature attribution for transparent security auditing.
Included an interactive dashboard for monitoring traffic, detected threats, and system performance.
Built a two-stage pipeline combining OCR (PaddleOCR, Tesseract), vision-language models (Qwen, Florence), and captioning (BLIP-2) for semantic meme understanding.
Designed context-aware generation with LLMs (LLaMA-3, GPT-4), with control over humor style, cultural framing, and multilingual adaptation beyond static template methods.
Research
Deep Learning Framework for Astronomical Image Deblurring
Vision Transformers, GANs, Computer Vision
Designed a physics-informed two-stage pipeline using a Vision Transformer for localized point spread function (PSF) estimation and DeblurGAN for deconvolution, mitigating atmospheric turbulence in large-scale sky surveys.
Presented at ISRO's S2-STEP 2025 and the 6th Indian Planetary Science Conference (IPSC 2025) at IIT Roorkee, selected from roughly 400 accepted papers out of 800+ submissions, with an oral presentation.
Sustainable Lunar Habitat Design Utilizing PEMFC and GPHS-RTG
Space systems
Researched hybrid power systems combining proton-exchange membrane fuel cells with radioisotope thermoelectric generators for long-duration lunar habitation.
Presented at ICPEH 2024 (PRL Ahmedabad) to 200+ researchers.