Likith

Built an AI-powered lip-sync platform enabling creators and enterprises to generate hyper-realistic talking-head videos from any face and audio track. Powered by MuseTalk v1.5 + Whisper for multi-language automatic speech recognition and perfect synchronization, running on GPU-accelerated cloud infrastructure.

• Architected end-to-end video processing pipeline on AWS (ECS/Fargate, S3, Lambda) supporting MP4, MOV, AVI, FLV, MKV formats up to 200 MB with 2–3 minute turnaround
• Integrated MuseTalk v1.5 model (VAE latent encoding, BiSeNet face blending) with Whisper for automatic language detection and audio phonetic analysis
• Implemented multi-language audio support with Whisper embeddings, automatically detecting language and optimizing sync parameters per language
• Designed secure, privacy-first architecture: automatic file deletion after processing, 7-day output retention, encrypted storage with strict PHI/PII isolation
• Built cloud processing layer optimizing GPU utilization and throughput for viral adoption—thousands of videos processed daily at sub-3-minute latency
• Deployed pre-built AI avatars (Jessica, Jack, Sarah) with one-click video generation, enabling creators to start without source footage
• Engineered robust error handling, graceful degradation, and user-friendly UI for seamless content creation workflows
• Currently available: 3 free videos per account (no card required); Pro plan with unlimited videos, multi-language batch processing, and API access (coming soon)

Implemented a decentralized IoT mesh communication network using the open-source Cluster Duck Protocol to simulate resilient connectivity during disaster scenarios.

• Deployed ESP32-based IoT nodes running the Cluster Duck firmware to form a LoRa-based mesh network with Wi-Fi client access
• Configured gateway nodes (Papa Duck architecture) to aggregate telemetry and forward messages to centralized endpoints via available backhaul connectivity
• Evaluated mesh formation behavior including node discovery, routing, and automatic rerouting during simulated node failures
• Tested network resilience under various failure scenarios to validate disaster communication capabilities

Decision-first task model where parent tasks hold deadlines while subtasks remain unscheduled prep steps—mirroring real-world prep before execution.

• Task structuring: only parents have date/time; subtasks are ordered, fully editable preparation steps to avoid artificial micromanagement.
• Suggested prep: deterministic keyword/pattern detection with curated templates; “system suggests structure, user owns the graph.”
• Decision engine: deterministic scoring of urgency, priority, readiness, remaining prep; outputs constrained Top-3/Top-5 actionable list to reduce decision fatigue.
• Dependency handling: blocked parents promote subtasks; conflicts surface warnings without auto-mutation.
• Execution layer: optional deep-focus sessions with timer + server persistence; streaks are derived from completion events (no manual input).
• Stack: Next.js + API routes + Prisma + SQLite; all business logic server-side for explainability with a clear path to layer AI later.

Outcome: an explainable, stable decision-intelligence system that optimizes for “what to do next,” not just task storage.

Reachy Mini humanoid assistant on Raspberry Pi 4 with CV + NLP for natural human-robot interaction.

• Voice + NLU pipeline for real-time speech understanding and responses.
• Gesture control algorithms for expressive arm/head movement and context-aware behaviors.
• Computer vision for object recognition/tracking driving interactive routines.
• Modular AI pipeline with customizable personality and response policies.
• Monitoring hooks into the multi-cloud dashboard for performance, thermals, and connectivity.

Impact: bridges digital intelligence with tangible interaction for practical robotics assistants.

Jetson Nano-based offline TTS pipeline optimized for privacy and sub-second inference.

• End-to-end GPU-accelerated inference with TensorFlow/PyTorch on CUDA.
• Custom-trained neural voices tuned for the target domain and edge constraints.
• Efficient training/evaluation loop plus thermal/resource monitoring.
• 100% offline operation—no cloud dependency for privacy-sensitive contexts.
• Integrated into multi-cloud dashboards to track inference latency, accuracy, and device health.

Impact: proves production-quality voice synthesis on embedded hardware for secure deployments.