Taneshq Gupta
Academic Record
Selected Work
- Architected a Vision-Language grounded segmentation model fusing a frozen CLIP ViT-B/32 text encoder with a LoRA-adapted DINOv2 ViT-S/14 backbone via a novel cross-attention grounding module, tuning only 3.19M of 88.3M parameters (3.6%).
- Designed a multi-scale Feature Pyramid Network and Dense-CSSE decoder fusing DenseNet-style feature reuse with concurrent spatial-channel squeeze-excitation for pixel-precise boundary delineation.
- Devised a two-model training strategy with 3× real-data oversampling, lifting real-world F1 from 0.8795 to 0.9546 on MASD and 0.8664 to 0.9336 on SBVPI, against a 0.9771 synthetic validation F1.
- Led a team of 10 for the SSBC 2026 Track 1 submission, training both regimes in under 3.5 hours on a single NVIDIA RTX A5000 24GB GPU.
- Collaborated with a cross-functional team of 8 peers to build an edge-AI system on an NVIDIA Jetson Nano using the MediaPipe Face Mesh pipeline.
- Implemented a unified attention score derived from EAR, MAR, and 3D Head Pose Estimation.
- Engineered a multi-modal alert engine: Visual (LCD), Audio (Voice), Haptic (Vibration).
- Built a real-time fleet dashboard with Flask & Redis, visualizing telemetry at 5s intervals.
- Tech Stack: Rust, SvelteKit, Postgres, TailwindCSS, TypeScript, Cloudinary, Railway, OpenStreetMap, Axum, Tokio.
- Installable as a PWA with PIN-code-based real-time map and 100+ category post filters; secure email auth with hashed passwords.
- Dynamic + persistent theme switching across 5 themes, with a user-friendly profile picture change protocol.
- Engineered a unified campus PWA consolidating 4 fragmented institutional systems (Governance, Grievances, Academics, Opportunities), designed to scale for 3,000+ students and faculty.
- Architected a highly concurrent, type-safe REST API using Rust and Axum delivering sub-50ms response times; secured 100% of endpoints via Google OAuth with strict RBAC across 4 user tiers.
- Designed a PostgreSQL schema with UUID primary keys and 7+ custom ENUM types; implemented a 4-stage automated grievance workflow with real-time status tracking and audit logging.
- Built a reactive, mobile-first frontend with Svelte and Vite, achieving a 100% PWA compliance score; cross-stack telemetry cut debugging time by 60%.
Recognition & Certifications
Awarded by IIT Mandi for reaching Grade A fluency at the Upper Intermediate (Level 4) tier of the institute's Japanese language program.
Official Japanese-Language Proficiency Test certification at the N5 level.
Tools & Technologies
Deep Learning
DINO CLIP Vision Transformers LoRA Cross-Attention FPN Squeeze-Excitation
Edge Computing & AI
SolidWorks (CAD) Jetson Nano/Xavier MediaPipe OpenCV TensorRT
Hosting & Domains
PWA / Service Workers Porkbun Railway Vercel Netlify Render Namecheap
Databases
PostgreSQL Redis (Time-series)
Languages
Python Rust C C++ Svelte Java SQL JavaScript TypeScript HTML/CSS
Libraries
PyTorch Pillow MediaPipe OpenStreetMap Axum Tokio Tailwind CSS pandas NumPy Matplotlib Chart.js