Edison
An AI agent that builds working circuits, and rebuilds a real breadboard from a photo.
Describe a circuit in plain English, and Edison's agent places every part and wire on a 3D breadboard through tool calls. A circuit simulator written from scratch then checks that the build works. Or photograph a real breadboard: a vision pipeline reads every part and lead, rebuilds the board on screen, and the agent explains what's wrong with it.
The three demos on the welcome card run instantly with no sign-in. Sign in with Google to ask your own questions (5 a day).
What it does
1. An agent that builds circuits
Ask for a circuit. The agent plans it, calls tools (place_battery, place_resistor, place_led, place_tl072, add_wire, …), and shows the build as a preview you accept in one click. Press Run and the simulator lights the LED.
2. It finds the mistakes on your real board, and fixes them
Edison simulates the board it rebuilt from your photo, so the agent can point at real problems: here, an op-amp turned around and two columns off, plus a missing V− wire. Say "fix it", accept the change, and the LED blinks.
| A student's real breadboard → what Edison builds from it |
|---|
|
3. A lab bench, not a toy
Bench supply, function generator, multimeter and scope; op-amps that clip at their rails; capacitors that charge over time. ENSC 220 labs come as step-by-step lab sheets that check each step against the simulation.
| Lab 2 with a mistake: the op-amp has no supply | Lab 2 running: a sine wave on the scope |
|---|
| |
How it works
flowchart LR
subgraph Agent["Agent (DeepSeek, tool calls)"]
P[Prompt + board as markdown] --> T[Tool calls<br/>place_* · add_wire · set_value]
end
T --> V[finishAIReply<br/>drop illegal actions,<br/>flag circuit problems]
V --> PR[Preview → Accept<br/>one undo step]
PR --> S[Simulator<br/>modified nodal analysis]
S --> R[LED on · currents · scope · mistakes]
R -. "what's wrong?" .-> P
- The agent. The board is sent as markdown (parts, holes, wires, simulation results), with a tool per part type drawn from the parts registry. Its reply is a list of actions.
finishAIReply drops anything illegal (occupied holes, wrong spans, unknown pins) and flags circuit problems. The page shows a preview, and Accept applies it as one undo step.
- The simulator. Modified nodal analysis, the method SPICE uses, written from scratch and running in the browser: Gaussian elimination per step; time-stepped transients for capacitors and sine sources; diodes, LEDs and zeners with their knees; op-amps that clip 1.5 V inside their rails.
- The eval.
npm run ai-eval sends 16 test prompts × 3 runs to the real model and grades each build in the simulator (does the LED light, does the op-amp output hit the expected voltage). The suite went from 24/48 to 35–38/48 as the prompts, tools and checks improved.
- Prompt goldens. Every prompt the agent sees is pinned in golden files, so a change to a part or a line of the prompt shows up as a reviewed diff, never a silent change.
Photo → breadboard
flowchart LR
A[Phone photo] --> B[Tap 4 corner holes<br/>homography onto the hole grid]
B --> C[Survey call<br/>Gemini Robotics-ER 2:<br/>parts, lead points, chip pin 1 + label]
C --> D[One batched zoom call<br/>≤12 crops, Set-of-Mark outlines]
D --> E[Snap leads to holes<br/>confirm screen: move any dot]
E --> F[PhotoImport<br/>legal board, same nets,<br/>chips seated from pin 1]
F --> G[Simulate + ask the agent]
- Geometry first. Four corner taps give a homography onto the exact hole grid, so the model is asked to point, not to count holes. Points snap to holes, and the model's own hole reading is only a tie-breaker.
- Few calls. A survey call returns every part with its lead points, plus each chip's pin-1 corner, notch and printed label. Then one batched call zooms on up to 12 parts at once, each crop outlined (Set-of-Mark). That's about 4 calls per photo instead of ~13, with no requests lost to rate limits.
- Chips are never guessed. A DIP op-amp (TL072, LM358, NE5532, …) is seated from its pin-1 and pin-4 points. If the notch disagrees or the points are only box corners, Edison asks ("which end is pin 1? ⇄ to turn it") instead of building it backwards. A 555 or 741 is flagged, not built as a TL072. Op-amp boards get a ±12 V bench supply, as in the real lab.
- Real handedness. Edison's board has the same handedness as a real breadboard, so a chip's pins land on the holes printed on your board. Boards printed the other way are flipped on import, with a warning.
Built with
| Layer | What |
|---|
| Frontend | Plain JavaScript and three.js (no framework, no build step): 3D breadboard, parts, wiring, HUD |
| Simulator | Modified nodal analysis, from scratch, in the browser |
| Agent | DeepSeek with tool calling (Gemini optional); a tool per part; validation before preview |
| Vision | Gemini Robotics-ER 2 (survey + batched zoom), homography in the browser |
| Backend | Node 20 with zero runtime dependencies: the AI routes, Google sign-in token checks (RS256 with node:crypto), daily limits |
| Hosting | Vercel (static app + one function), Firebase Google sign-in, Upstash Redis for daily counters |
| Tests | ~3,400 unit tests (Vitest) and ~360 browser tests (Playwright); no test ever calls a real AI |
Public demo safety: the three demos replay saved answers, so they cost nothing. Live questions need Google sign-in, and each user gets 5 a day. A global daily cap keeps the bill bounded.
How it was built
Edison is built with an agentic workflow. Every change goes test-writer → builder → reviewer, as separate Claude Code agents. Hooks enforce their lanes: the test-writer can only touch tests, and the builder can't touch them. Each task's tests must fail first, and a change ships to dev only after lint, unit and browser tests pass. The details are in docs/CONTRIBUTING.md; the design docs are in docs/.
Run it locally
cd Plugged && npm ci
# Plugged/backend/.env: DEEPSEEK_API_KEY=…, GEMINI_API_KEY=…, AI_PROVIDER=deepseek
cd backend && node server.js
# open http://localhost:5001 (no keys? AI_PROVIDER=fixture replays recorded answers)
npm run check (lint) · npm test (unit) · npm run e2e (browser)
Team
Built at StormHacks 2026 by Aarmen Sidhu (all code: agent, simulator, vision pipeline, deploy), with Bhavan Thandi (circuit design and testing on real parts) and Jessi Sidhu. Mentored by Raymond Shen.
