Projects_
Vulcan: Self-Driving Campus EV (Jan 2023 – present · team lead)
A golf cart converted into an autonomous campus vehicle, a VIT-AP initiative. I lead perception and navigation software. I trained DeepLabV3+ on Mapillary for road-scene segmentation (67% multi-class accuracy), wrote the lane-following and turn-handling logic, and built an end-to-end driving pipeline in CARLA using Soft Actor-Critic, fine-tuned with DAgger. It runs in real time on an NVIDIA Jetson, fusing camera, LiDAR and radar.
ROS · CARLA · PyTorch · DeepLabV3+ · SAC · DAgger · Mapillary Vistas · NVIDIA Jetson
16fps: Multi-Video Generation Agent (Oct 2025 – present · self-employed, SaaS)
An agentic system that turns a text prompt into a consistent video of a minute or more. Several agents plan over a long context and keep memory. They call tools and use reflection to catch and fix their own mistakes, which keeps characters, scenes and story consistent across 60+ second clips at 16 FPS. It renders several videos in parallel with style transfer and audio sync, and serves them from zero-downtime cloud infrastructure. Commercial rights and operational access were sold to a media production company.
Python · LangChain · Stable Video Diffusion · AnimateDiff · GCP
EchoSight: Assistive Navigation Glasses (Dec 2024 – present · patent pending)
Wearable glasses that help visually impaired people get around. They detect objects in real time and give spoken feedback through Bluetooth earbuds. YOLOv8 is fine-tuned on 50+ custom classes and runs at 45+ FPS on a Raspberry Pi. Monocular depth estimation gives direction and distance, and alerts adapt to the context and location. Awardee at the Vikas 2024 Innovation Challenge.
YOLOv8 · Roboflow · Raspberry Pi · OpenCV · GCP · TTS
WOAT: Context-Aware Conversational Agent (Jul – Nov 2023)
An agent that reads past chats, infers each relationship and tone, and drafts WhatsApp replies in the user's own style. The integration layer (Python + Node.js) handles context memory, contact-aware behaviour, QR login and fallback modes. Separate modules cover tone adaptation, understanding the situation, priority ranking and adversarial safety testing.
Gemini API · GCP · Node.js · Python
DeepShield: Face-Swap Deepfake Detection (Jul – Nov 2023 · iSAI 2025)
A CNN deepfake detector: +22% classification accuracy and 18% fewer false positives. A synthetic-video generator adds 40% more variety to the training data, and adversarial stress tests check how robust it is. Published at the 20th International Conference, iSAI 2025. cite →
NeuroSecAPI: Neuro-Symbolic API Vulnerability Detection (AIR Centre, 2024–2025)
Symbolic reasoning combined with LLM-generated payloads (SQLi, XSS, BOLA) and adaptive fuzzing. +34% vulnerabilities found compared with SAST/DAST, and 70% less manual testing. A modular, multi-service pipeline with OWASP API Top 10 risk scoring.
Flask · OWASP · LLMs · neuro-symbolic AI
ScreenAutomate: Gesture + Voice Control (Oct 2023 – Jan 2024)
Control a display with hand gestures and voice commands, active whenever you're within Bluetooth range. Built for accessibility and remote desktops.
More photos (events, awards, certificates) are in the gallery.