I build the interfaces AI products need but rarely have.

AI operations platform

Agents that act on a seller's real revenue can't ship their output straight to the customer — but a review queue that slows every job to human speed defeats the point.

I own front-end delivery: the agent run surfaces for Analyst, Risk Manager and Creative Studio, SSE streaming that holds up across jobs measured in minutes, and a reusable role-config layer that routes AI-generated copy and images through four levels of sign-off. Also the Amazon SP-API and Ads OAuth integration, and the state architecture underneath — Zustand for client state, TanStack Query for server state.

• Next.js
• TypeScript
• Zustand
• TanStack Query
• Tailwind
• SSE

AI lead enrichment platform

Enrichment runs bill per record, so an unbounded pipeline turns a prospecting list into an open invoice.

A NestJS and BullMQ pipeline on Redis that scrapes, enriches and scores Amazon seller data, with deduplication, stop conditions and per-job API cost tracking. Next.js dashboard streaming job progress over SSE. Structured leads land in PostgreSQL as JSONB; raw payloads stay in MongoDB Atlas.

• Next.js
• NestJS
• PostgreSQL
• MongoDB
• Redis
• BullMQ
• OpenAI

24-hour hackathon build

Triage needs to know who is messaging and what came before — signal a message body alone doesn't carry.

An agent that sorts WhatsApp messages into notify, digest or mute, reading chat screenshots as multimodal input alongside CSV context for sender priority and history. Most of the work was in prompt structure, tuned specifically to cut false-notify cases.

• TypeScript
• Node
• multimodal LLM prompting