Projects

Built a fully on-device conversational voicebot with end-to-end response latency under 1.5 seconds, designed to run without a GPU or cloud connection. The system is a cascaded architecture chaining three independently optimized low-latency modules: speech-to-text (STT) for transcription, a compact large language model (LLM) for reasoning and response generation, and text-to-speech (TTS) for natural voice output. Each stage is selected and tuned for minimal inference time — quantized models, streaming and chunked processing, and pipelined execution so modules overlap rather than run strictly in sequence. The result deploys on commodity hardware such as laptop CPUs and mobile phones, keeping conversations private, offline-capable, and responsive.

Proposed a sequence-to-sequence masked transformer (inspired by the Masked Trajectory Transformer and BetaAlign) that generalizes to unseen gene families. Achieved higher alignment accuracy than classic tools such as ClustalW and MAFFT, with up to 13% runtime reduction.

Industrial project with CVRDE, DRDO. Fine-tuned GANs to disentangle latent factors (shape, size, texture, color) for precise control and clean foreground and background separation. Added RGB to Thermal-IR translation via edge-guided GANs, improving foreground fidelity and yielding a 6% accuracy lift on a custom set.

Built an end-to-end YOLOv5 pipeline for helmet, triple-riding, and overspeed detection; achieved F1 = 0.79 and 89% deployment accuracy on Jetson Nano. Incubated at Nirmaan and integrated with existing CCTV infrastructure for real-time tracking.