Neural Speak: AI Voice SaaS
Python TypeScript Next.js React FastAPI PostgreSQL AWS
Production text-to-speech SaaS that converts user scripts into studio-quality speech, with multi-language output and voice cloning. Built credits-based billing and subscription flows with a customer portal for invoices and usage. Delivered end-to-end audio project management with secure authentication and scalable deployment.
Mechanistic Interpretability for Steering VLA Models
Python PyTorch OpenVLA Sparse Autoencoders LIBERO
Reproduced mechanistic activation steering on OpenVLA and pi0.5 to induce targeted behavior changes without retraining. Evaluated controllability with extensive LIBERO rollouts and compared token-space interventions against sparse autoencoder probes. Trained SAEs on ~10M activations to analyze task-manifold geometry, feature disentanglement, and steering reliability across tasks.
Self-Destructing Real-Time Chat
Next.js TypeScript Redis Elysia.js Upstash Realtime WebSockets
Built privacy-focused real-time chat application with automatic room destruction, strict two-user access limits, and complete message wiping from Redis using TTLs. Implemented type-safe backend with Elysia.js and real-time WebSocket messaging via Upstash Realtime, enforcing access control using HTTP-only cookie tokens.
Diffusion Image Generation and Editing*
Python PyTorch
Built diffusion pipelines for image synthesis and editing with guided sampling, inpainting, and SDEdit-style image-to-image translation. Implemented the forward noising and iterative denoising loops with classifier-free guidance to study fidelity and prompt adherence. Added a custom UNet baseline for controllable generation experiments.
Automatic Panorama Stitching and Seam Blending*
Python
Built an end-to-end panorama tool that detects and matches features, estimates homographies, and warps images into a shared canvas. Implemented inverse warping with bilinear sampling to avoid holes and reduce aliasing. Added feathered blending and gain compensation to hide seams and exposure shifts for clean stitched outputs.
*Source available on request