Skip to content
pavel·smolin
enrupl

Staff Product Engineer · AI Systems

Building production AI systems.

I’m a Staff-level product engineer with 10+ years of experience building and shipping production software across AI, fintech, web, mobile, and Web3. I work across architecture, TypeScript and Node.js backends, React and Next.js, React Native, LLM systems, integrations, and production reliability.

Work with meView my work
Pavel Smolin overlooking the Golden Gate Bridge in San Francisco
San Francisco · 2026
10+years shipping production software
Web + mobileCouncil, an AI product built and shipped solo
82%of AI runs no longer wait on a planner call

What I do

The problems I’m most useful on sit where product decisions, architecture, and hands-on delivery meet.

Complex products, from idea to production.

01

AI Systems

Design and ship production LLM systems: multi-model orchestration, structured outputs, streaming, evaluation, reliability, cost optimization, and agentic workflows.

02

Product Engineering

Take products from idea to production across web and mobile, combining product judgment with hands-on engineering.

03

Architecture & Platforms

Design maintainable TypeScript systems, APIs, integrations, infrastructure, and developer workflows that can evolve as the product grows.

Case studies

Systems I owned end to end: the problem, the decisions, and what changed.

Evidence, not a portfolio.

AI product · Web, iOS, Android

Council: a multi-model AI product, built and shipped solo

  • 3models per session
  • 2 → 1payment providers, one source of truth
  • 1backend behind web and mobile
Problem
One model gives you one opinion. Council asks several independent models the same question and reconciles their answers into a single recommendation with an explicit consensus score.
What I owned
All of it: product, architecture, the Next.js and Supabase backend, the web app, the Expo mobile app, subscriptions, and release.
Architecture
Parallel fan-out across LLM providers, streamed to the client as Server-Sent Events. Every provider response is validated into a typed shape, and a failing model degrades the run instead of failing it. Stripe and RevenueCat are reconciled into one subscription table.
Outcome
Live on the web, on the App Store and on Google Play, with one backend serving both clients.
Read the case studytrycouncil.app App Store Google Play

AI reliability · Latency

A 300 ms gate in front of a 15-second AI pipeline

  • 82%of runs skip the planner wait
  • ~300 msmedian gate latency
  • 0.94 / 1.00precision / recall on the eval set
Problem
Every run waited about a second for a research planner that answered “no” 82% of the time, and inputs that weren’t questions still paid for four model calls.
What I owned
The diagnosis from production data, the gate design, an evaluation set with a pass bar, and the rollout.
Architecture
A typed judgment call returns probabilities, not text, and code applies the thresholds. Every failure path fails open, so the gate can never break the pipeline behind it.
Outcome
The planner wait is gone on 82% of runs, and junk input is refused before it reaches three frontier models.
Read the case study

Web performance · SEO

This site: a perfect Lighthouse score on a real page

  • 4 × 100Lighthouse categories
  • 3.5 MB → <200 KBfor the four heaviest images
Problem
A trilingual site with photography and long-form writing, not a stripped-down demo, that still had to be fast, accessible, and indexable.
What I owned
Design system, internationalization, structured data, deployment, and the performance work.
Architecture
Static Astro build, responsive AVIF and WebP images with reserved dimensions, hreflang alternates, a sitemap with real modification dates, and JSON-LD on every post.
Outcome
100 across Performance, Accessibility, Best Practices, and SEO on PageSpeed mobile.
Read the case study

Writing

Notes from production systems.

View all articles
Guide4 min read

How to Add the Supabase MCP Server to Claude Code

One command connects Supabase to Claude Code through the hosted MCP server: OAuth or a personal access token, project scoping, read-only mode, and the tools to allowlist.

Experience

Built in the messy middle between strategy and shipping.

Ten years across early-stage products, scaling platforms, and technically demanding domains: AI, fintech, and Web3.

  1. Now

    Staff-level product engineer & consultant

    Building and operating production AI systems end to end, and helping teams with LLM architecture, reliability, latency, and cost.

  2. Recent

    Technical leadership

    Architecture and delivery across TypeScript platforms, fintech, and Web3 products, as the engineer who owns the hard parts.

  3. Foundation

    Senior software engineer

    Web and mobile products built end to end: interfaces, APIs, data models, and infrastructure.

Toolkit

Tools that support the work.

Chosen for the product, team, and stage, not the other way around.

AI
  • OpenAI
  • Anthropic
  • Gemini
  • LLM orchestration
  • Structured outputs
  • Evaluation
Product
  • TypeScript
  • React
  • Next.js
  • React Native
  • Expo
Backend
  • Node.js
  • NestJS
  • Fastify
  • GraphQL
  • Supabase
  • PostgreSQL
Infrastructure
  • Vercel
  • AWS
  • Cloudflare
  • EAS
  • PostHog
Web3
  • NEAR
  • Ethereum
  • Stellar
  • Wallet integrations

Have a difficult product or engineering problem?

I work with teams building AI products, developer platforms, fintech, and other technically demanding software.

Discuss a project