CircuitMind.AI

Landing Page

Inspiration

Every hardware startup starts with an engineer sketching circuits on a whiteboard, then spending days translating that into a real design. Software engineers have Copilot — hardware engineers have nothing.

When Gemini 3 launched with multimodal reasoning and structured output, we saw an unprecedented opportunity: an AI that looks at a hand-drawn circuit sketch and produces a validated hardware design in 60 seconds. Not a chatbot. An autonomous design agent.

What it does

CircuitMind.AI takes 3 types of input — a hand-drawn sketch photo, a one-sentence description, or manual module selection — and runs a 4-step agentic pipeline:

Step	Action
Perceive	Extracts electrical parameters, identifies ambiguities
Generate	Creates 3 differentiated solutions (JSON Schema-enforced)
Validate	AI self-reviews for voltage mismatches, bus conflicts, missing components
Iterate	Auto-fixes critical issues — no human intervention needed

Each project produces: L1 architecture diagrams, R&D swim-lane workflows, BOM estimates, milestones, and open engineering questions.

How we built it

4+ Gemini 3 API calls per project orchestrated through runAgenticPipeline() in src/lib/gemini.ts
Multimodal sketch analysis via inlineData (base64 images) sent to Gemini 3 Vision
JSON Schema enforcement (responseSchema with 250+ fields) — zero parse failures
Pro + Flash split: Flash for generation, Pro for validation, with automatic fallback
Stack: React 19, TypeScript strict, Vite 7, Tailwind CSS, @xyflow/react, react-i18next

Challenges we ran into

Gemini 3 Pro free quota = 0 — built automatic Pro→Flash fallback; pipeline still runs 4 independent calls
JSON reliability — prompt-based output failed ~15%; switching to native responseSchema dropped failures to zero
Avoiding "prompt wrapper" — redesigned from single-call to true multi-step agent with conditional branching

Accomplishments that we're proud of

True agentic AI: 4-step self-validating pipeline, not a wrapper
Zero JSON parse failures with 250+ field schema enforcement
Sketch-to-design in seconds: napkin photo → complete circuit specification
Full product: landing page, dashboard, project management, case library, component DB, bilingual UI

What we learned

responseSchema is Gemini 3's killer feature. It turns "hope the AI returns valid JSON" into guaranteed type-safe output. The validate→iterate loop caught real engineering errors (voltage mismatches, missing pull-ups) that generation alone missed. Multimodal vision for domain-specific analysis opens entirely new application categories.