The same buck, built by AI in Autodesk Fusion 360

Open the 3D model full screen

This is the first video on YouTube of an AI designing a complete circuit board in Autodesk Fusion 360. AI PCB videos are starting to appear for other tools, but the AI-in-Fusion videos out there are all mechanical CAD. To make it possible we built our own MCP server on steroids, the Fusion Bridge: it lets an AI automate Fusion end to end for electronics, from signing in and building the part library to the schematic, the board, the copper and the fab files.

Here it builds the same 12 V to 5 V, 1 A TPS54202 buck as the KiCad build on the main page, from one sentence, as a complete board: Claude on Adom AI Flow inside Adom Hydrogen, with the Fusion Bridge driving Fusion through direct commands instead of clicks. A routed, poured, analysed board with Gerbers, a BOM and a pick-and-place file came out.

The videos

Watch on YouTube

1st Ever: AI Designs a Complete PCB in Fusion 360 (4 min 47 s). This run: the same 12 V to 5 V buck as part 1, built from one prompt as a complete board in Autodesk Fusion 360. Claude on Adom AI Flow pulls the parts from the Adom wiki, and the Fusion Bridge signs in to Fusion, builds the library, draws the schematic and lays out the board. Along the way AI Flow sends Claude back to fix its own placement and heatsink vias, Adom Fields checks the current and heat, and the run ends with Gerbers, a BOM and a pick-and-place file. It took 2 h 42 min from prompt to finished board.

Watch on YouTube

Part 1: AI PCB Design for Engineers (about 4 min). The first build of this board, in KiCad 10, and the precursor to this Fusion 360 run. It covers the prompt in Adom Hydrogen, the circuit, three SPICE engines in agreement, the live layout, the current and heat analysis and the rework it caused. The same brief built in Altium is planned as the next part. The KiCad build.

Parts

Every part on this board has its own page on the Adom wiki, with its symbol, footprint and 3D model, ready to reuse in your own designs. Click a part to open its page. These are the same parts as the KiCad build, whose BOM lists every distributor number; the Fusion board has six edge contacts where the KiCad one has seven.

part	qty	refs	value	what it does	source
TPS54202DDCR Texas Instruments	1	U1	TPS54202DDCR	sync buck 4.5-28 V 2 A 500 kHz	Mouser 595-TPS54202DDCR
AMPLH5030S-150MT Abracon	1	L1	15uH	output inductor	Mouser 815-AMPLH5030S-150T
EEE-FK1V470P Panasonic	1	C1	47uF 35V	input bulk (cable damping)	Mouser 667-EEE-FK1V470P
CC1206KKX7R8BB106 Yageo	2	C2, C3	10uF 25V X7R	input ceramic	Mouser 603-CC126KKX7R8BB106
0603B104J500CT Walsin	2	C4, C5	100nF 50V X7R	VIN pin HF bypass	Adom stock
CL21A226MPQNNNE Samsung	3	C6, C7, C8	22uF 10V X5R	output ceramic	Adom stock
0402N680J500CT Walsin	1	C9	68pF C0G	feed-forward across Rtop	Adom stock
CR0402-FX-6802GLF Bourns	1	R1	68k	FB top (series part 1)	Adom stock
CR0402-FX-5601GLF Bourns	1	R2	5.6k	FB top (series part 2)	Adom stock
CR0402-FX-1002GLF Bourns	3	R3, R7, R8	10k	FB bottom	Adom stock
CR0402-FX-1004GLF Bourns	1	R4	1M	EN UVLO top	Adom stock
CR0402-FX-1503GLF Bourns	1	R5	150k	EN UVLO bottom	Adom stock
CR0402-FX-2201GLF Bourns	1	R6	2.2k	LED resistor	Adom stock
IN-S63BTG Inolux	1	D1	green LED	5 V power-good indicator	Adom stock
MachineContactMedium Adom	6	J1 to J6	MachineContactMedium	molecule edge contact	Adom stock
MachinePinMediumShort Adom	4	MP1 to MP4	MachinePinMediumShort	corner machine pin (GND)	Adom stock

Where the current and heat go

Watch on YouTube

Where the Current and Heat Go (8 min 44 s). The KiCad build of this board solved in Adom Fields, in eleven narrated views: current density, heat flow, temperature, the current and voltage flowing, energy, how fast a switch-on travels, the energy cycle, charge, the magnetic field and EMC. Fields is the same solver that checked the Fusion board's copper in this run. The numbers.

At a glance

The prompt	"Use adom-aiflow to design a 12 V to 5 V, 1 A buck converter using the TPS54202, and build it as a complete board in Autodesk Fusion."
The AI	Claude Code on Claude Opus 5.5, in Hydrogen
Prompt to finished board	2 h 42 min, 21 steps, 4 loop-backs, none skipped
Design check	0 errors, 0 unconnected
Current	peak 53 A/mm² across the copper, 0 issues
Heat	U1 rises 49 °C in still air (28.5 °C with airflow)
Board	28 x 20 mm Adom molecule, 2 layers
Out of the run	Gerbers, BOM and pick-and-place (CPL)

The tools

tool	what it did here
Adom Hydrogen	The workbench: VS Code and Claude on the left, web views (timeline, wiki, SPICE, Fields) on the right.
Adom AI Flow	The run itself: plans the 21 steps, checks each one, and sends Claude back when a step fails its check.
Fusion Bridge	Drives Fusion 360 on a Windows PC from the cloud: sign-in, library, schematic, board, exports.
Adom wiki component pages	Where every part comes from: symbol, footprint, a checked 3D model and distributor numbers.
Adom SPICE	The dashboard: ngspice in the container, with LTspice and PSpice for TI one click away, all on one set of axes.
Adom Fields	Current density and heat on the real copper.

The run, step by step

1. One prompt

The video opens on the whole job in one sentence, typed into Claude Code inside Adom Hydrogen. Naming adom-aiflow is what turns the request into a measured run with a clock, a plan and checks, instead of an open-ended chat. Hydrogen puts VS Code and Claude on the left and live web views on the right, in one window, so everything the run produces shows up beside the conversation that produced it. The AI works in a cloud container, yet it builds the board in Fusion on a real Windows PC through Adom Bridge.

Try it yourself. Copy the prompt and paste it into Claude Code in Adom Hydrogen:

2. The AI Flow timeline

AI Flow lays the job out as 21 steps, from intake to the finish line, and shows them as a live timeline in one of Hydrogen's web views. Every step has to pass its own check before the next one starts, and the clock runs from the prompt to the last export. Each arc on the left is a loop-back: a later step found a problem and sent Claude back to an earlier one, and this run had four. Because the timeline is a page in the same window, you watch the plan, the current step and every return as they happen, with no switching between apps. (AI Flow)

3. Parts from the Adom wiki

Every part on the board comes from its own component page on the Adom wiki, like the TPS54202DDCR here. Each page carries a symbol, a footprint, a 3D model checked Z-up, and Mouser, DigiKey and LCSC numbers, so the AI starts from parts that were already verified instead of guessing at CAD. Hydrogen opens these pages in a web view next to the chat, so the part Claude is reasoning about is on screen while it does. Parts that are not on the wiki yet are fetched and built by Chip Fetcher.

The green power LED, IN-S63BTG, shows a wiki feature you won't find on other part sites: a states widget, with one 3D model per state, off and lit. A viewer can show the board with the LED on, and a probing simulation can check that the LED should be lit at that test step. The widget is one of the reusable component widgets any part page can carry. Flipping it from Off to On in the video is a live view of the model, not a picture.

4. The bridge signs in to Fusion

The PC for this step was not signed in to Fusion yet, so the Fusion Bridge handled the sign-in itself: it pressed Fusion's Sign in button and drove the Autodesk page in the browser. It never types a password or a 2FA code; those stay with the person. This is the part most automation stops at, because a signed-out app is a dead end for an AI. Hydrogen and Adom Bridge reach a whole desktop, not just one program, so the AI can get past the sign-in and keep going.

Fusion then said the license seat was still in use on another PC. The bridge chose to suspend Fusion there and continue here, the decision a person would make, and carried on. Hydrogen drives several Windows PCs as one mesh through Adom Bridge, so moving the seat between machines is a normal part of a run, not an interruption.

5. One project library

AI Flow builds one Fusion library for the project, with a symbol, a footprint and a bound 3D model for every part. Here is the TPS54202 symbol, drawn with its chip outline so the schematic reads like the real package. The Fusion Bridge creates the library through Fusion's own API and commands, with no library dialogs to click through. A Fusion parts library is normally hours of careful hand work; here it is one step of the run. (Fusion Bridge)

The footprint comes from the same wiki page as the symbol, so pads, pin numbers and the 3D model agree by construction. Every part gets the same treatment, down to the 0.5 mm probe test pad that the probing workcell lands on. Because the library is built from verified wiki data, the board that comes out of Fusion matches the parts you would order.

6. The schematic

Claude draws the schematic live in Fusion, in labelled groups. Every pin is checked against the netlist the circuit step produced, so a wrong connection fails the step instead of reaching the board. The Fusion Bridge places and wires parts through Fusion's own electronics commands, so the drawing is real Fusion data you can open and edit, not an imported picture. Hydrogen and the bridge let the AI work Fusion's electronics editor the way a person would, but by command, a whole schematic in minutes.

7. SPICE

Before any copper, the circuit is simulated in ngspice, running right in the container: the soft start from 12 V, a load step from 0.5 A to 1 A, and the inductor current through both. The dashboard can also run the same circuit in LTspice or in PSpice for TI on a Windows PC, one click each, to cross-check the result. Hydrogen keeps the plots in a web view next to the design, so the numbers that justify each part value sit on the same screen as the circuit. (Adom SPICE, installed with the SPICE skill pack)

8. Placement, and the first loop-back

Claude places the parts, then AI Flow checks the result against the layout rules before routing starts. Here the check found the switch-node trace detouring 18.9 mm because of where C5 sat, so the timeline drew an arc and sent Claude back to placement. The glowing ring marks C5 in its first spot. This is the core of AI Flow: the AI is graded step by step, and a failed check sends it back instead of letting a bad board through.

C5 moved next to the pins it serves, and the switch-node loop got short. A second loop-back came from routing: the VOUT feed passed 0.15 mm from R6's LED pad, so Claude turned R6 to face the feed. Each return is recorded on the timeline with its reason, so you can audit every decision the AI made, not just the final board.

9. Routing, heatsink vias and pours

The board is routed to 100 percent, then the heatsink vias, the copper pours and the rounded corners each land as their own checked pass. A review of the vias found U1 had only 1 of its 6 planned heatsink vias, and AI Flow sent Claude back to routing: clear the copper, re-route around U1, place the full array and pour again. The six vias now tie U1's ground pin to the bottom ground plane, all within 2.75 mm. Every copper change happens in Fusion through the bridge's commands, so the result is a normal Fusion board design. (AI Flow)

The video then shows each pour net in turn, highlighted on its own. GND fills most of the board and gives the regulator's heat somewhere to go. Walking the nets one at a time is how an engineer checks a pour, and here the AI does it as a step of the run.

The switch node gets a small, solid SW patch, kept short on purpose because it is the noisiest net on a buck. A large SW pour would radiate; a tight one keeps the hot loop small. That judgement comes from AI Flow's power-layout rules, which the run follows and checks.

10. Current and heat

Adom Fields solves the real copper of the finished board for current density and temperature, and shows it in 3D in a Hydrogen web view. The peak current density is 53 A/mm², and the heatsink vias carry about a third of U1's 0.25 W into the ground plane. This is a field solve on the real layout, not a rule of thumb. The analysis is a step of the run, so a result over budget sends Claude back to fix the copper instead of just being reported.

11. Silkscreen

The silkscreen is laid down one item at a time: framed tags on every edge contact, matched frames on the corner pins, and each value in small text beside its reference. Text heights follow the fab's minimums, so every label prints. The rules that came out of this run, such as frame sizes, equal widths and minimum text heights per fab, are now part of AI Flow's silkscreen skill for every future board. Good silkscreen is what makes a board easy to assemble, probe and debug.

12. The board in 3D

Fusion's 3D view shows every part sitting flat on its pads, which is the check that the library's 3D models were bound correctly. Each model came from its wiki page, already checked for orientation. The same board is in the 3D viewer at the top of this page, exported from Fusion and running in the wiki's own viewer.

13. The finish line

The timeline reaches 21 of 21 steps: the finish line is Gerbers, a BOM and a pick-and-place file exported from Fusion. AI Flow only calls the run finished when every step has passed its check, so the board is routed, design-rule clean, poured, analysed and exported. The clock stops at 2 h 42 min from the prompt. What you get is a real Fusion project plus the files a fab needs.

What it cost

One run on Claude Opus 5.5, prompt to finished board, 2 h 42 min. Measured from the run's own session log and Claude's plan usage meter:

plan	cost of this run
API (pay per token)	about $24
Max 20x ($200/month)	about 4 % of a week, about $1.85
Max 5x ($100/month)	about 16 % of a week, about $3.70
Pro ($20/month)	about 80 % of a week, about $3.70, with waits for the limit to reset

The API figure is 203 k output tokens, 83 M cached input tokens read and 453 k written, at Opus 5.5 list prices.

After the run

• Silkscreen: the pass shown in step 11 was finished after the run, one tag at a time.
• Test points: TP5 came off the board; its ground probe moved to J5 pin 1.
• Final check: 0 errors and 0 airwires, saved as the final cloud version, and the fab files were exported again.
• Hottest spot: 57 °C rise in still air against a 55 °C budget. A 28 x 20 mm board shedding 0.37 W into still air is limited by its size, not its layout.
• Sourcing: the TPS54202 was out of stock at Mouser and DigiKey when this was built (140-day lead), so order it early.

How the video was made

• Capture: the whole run was recorded on the Windows desktop at 2560 x 1440 with Adom Bridge's built-in recorder, about 2 h 50 min of footage.
• More than one PC: Hydrogen drives several Windows PCs as one mesh. The run happened on one. The opening prompt, the sign-in, the library and the wiki pages were re-shot afterwards, some on other PCs, so the main take stayed untouched.
• The cut: a script cuts about 50 shots from the footage, sizes each one to its narration line, and renders every zoom with smooth sub-pixel motion. The rings and arrows that point at C5 and R6 are drawn on top in brand teal.
• The cost scene: a slow walk around the finished board, rendered natively in Fusion, behind the numbers above.
• The thumbnail: the board is a render from Fusion, placed over a screenshot of the same board in Fusion's 2D editor.