Building AI products

Magnetic

Co-founded an AI marketing agent for Shopify sellers. Led product and engineering from ideation to MVP with a team of 5.

Agent vs UI balance

Started agent-first — AI drives everything. Users didn't want that for structured tasks. We landed on a per-screen balance: some screens need predictable UI with agent as backup, others work best as conversation.

UI leads

Onboarding

Setup

Account connection

Clear steps, predictable flow. Agent helps when things go wrong.

Agent leads

Creative generation

Campaign building

Reporting

Merchant describes what they want, AI handles execution.

Studios

One agent trying to handle images, audiences, and campaigns at once = more hallucinations and wrong actions. We split the agent into “studios” — each is a scoped context with its own instructions, tools, and UI, but shared conversation history. The agent only sees what's relevant to the current task.

Images

Generate & edit creatives

Image gen, product photos, variations

Targeting

Audience & demographics

Demographics, interests, lookalikes

Campaign

Budget, schedule, publish

Meta Ads API, budget rules, preview

Each studio filters its tools and injects only relevant state — smaller context = fewer hallucinations

Technical decisions

Model selection

Cerebras GLM-4

Tested Groq, Claude, Gemini. Needed fast enough for good UX, smart enough for reliable tool calls. Model choice directly impacts unit economics.

Action confirmation

Per action type

Low-stakes actions (image variations, copy) flow naturally. High-stakes (budgets, publish) always require explicit user confirmation.

Context grounding

Readable prose, scoped

UI state injected as human-readable text, not JSON. State builders extract only studio-relevant fields to keep context small.

Previously: Shipped AI Formulas in Structure — used AI to remove the technical barrier from our competitive moat and drive adoption. Different problem, simpler tech, same focus on business results.

← MS Planner to Jira Home