What I'm working on.

Building

Rasphia — AI-native conversational commerce platform using RAG + multimodal models (OpenAI, Claude, Gemini) for cross-site product search and purchase. Currently expanding the product catalogue coverage and improving cross-site retrieval accuracy. Live at rasphia.com.

SeedhaPe — UPI middleware for zero-fee merchant collections. Verifies customer payments made directly to merchant bank accounts, then orchestrates trusted post-payment fulfillment without acting as a money-moving processor. Focused on expanding the merchant onboarding flow. Live at seedhape.com.

Consulting

Taking on select consulting engagements with US-based service businesses and brands around AI discoverability, conversational commerce, and agentic payment architecture. The three core offerings are on the homepage.

If you have a specific bottleneck in mind, the fastest path is a short scoping call.

Thinking about

Agentic payment governance. The payment rails (x402, Stripe, Visa/Mastercard agent specs) are mostly solved. The open problem is permission models — spending envelopes, counterparty allowlists, audit trails. I wrote about this in a recent post.

LLM evaluation frameworks at production scale. Most eval setups I see are tested against synthetic data and break on real customer inputs within two weeks of deploy. I'm working on a lightweight harness design that holds up under distribution shift.

Verifiable compute + ML convergence. Running ML inside RISC-V environments (the work I did at Cartesi) is getting more tractable as toolchains mature. Worth watching whether this opens a path to auditable inference for regulated industries.

Reading

Anthropic's interpretability research — particularly the work on sparse autoencoders and feature geometry. The x402 protocol spec and the Visa/Mastercard agentic payment standards. Gwern's writing on forecasting and rationality, which I return to periodically.