AI Flow: Reverse Engineering

1
00:00:01,958 --> 00:00:06,562
This is a three phase motor controller. A power connector, a bank of

2
00:00:06,582 --> 00:00:08,193
capacitors, six MOSFETs.

3
00:00:08,213 --> 00:00:13,284
A gate driver and a microcontroller in the middle, each turned 45 degrees.

4
00:00:13,304 --> 00:00:16,653
We have no design files in this flow. Just four photos.

5
00:00:17,216 --> 00:00:20,418
Turn it over, and the copper shows through the solder mask.

6
00:00:20,438 --> 00:00:25,275
Pours, fine tracks, gaps between them, and more than a thousand tiny domes

7
00:00:25,295 --> 00:00:27,638
where vias push up through the mask.

8
00:00:27,658 --> 00:00:30,706
Everything from here on is read from pictures like these.

9
00:00:31,289 --> 00:00:35,729
The job: find the parts and the blocks. Copy all the copper. Work out what

10
00:00:35,749 --> 00:00:40,429
can't be seen, propose a layer count, say how sure you are, and list the tests

11
00:00:40,449 --> 00:00:42,057
that would settle the rest.

12
00:00:42,077 --> 00:00:46,908
Then a harder bar. Put every piece of copper on a net. Make the layers match

13
00:00:46,928 --> 00:00:51,887
the photos. And keep going until confidence measures 90%, or say why it can't.

14
00:00:52,459 --> 00:00:57,078
This is AI Flow: Reverse Engineering. A Rust program does the measuring, and

15
00:00:57,098 --> 00:01:00,497
an AI supplies the judgment, like what a part is called.

16
00:01:00,517 --> 00:01:05,453
11 steps, from intake to report. Each one has a check that has to pass, and

17
00:01:05,473 --> 00:01:06,973
each one gets recorded.

18
00:01:07,592 --> 00:01:12,025
First, the photos. It finds the board edge in each one and straightens it.

19
00:01:12,045 --> 00:01:16,486
The bottom gets turned over and pinned to the top through 40 holes that both

20
00:01:16,506 --> 00:01:17,191
sides share.

21
00:01:17,211 --> 00:01:22,108
With one hole held out each time, they land within about 0.03 mm.

22
00:01:22,682 --> 00:01:28,538
Scale comes from the board itself. Header rows have standard pitches, 2 mm and

23
00:01:28,558 --> 00:01:29,140
2.54 mm.

24
00:01:29,160 --> 00:01:34,695
The driver chip's 6 mm body confirms which standard they are. That gives 46.6

25
00:01:34,715 --> 00:01:36,281
pixels per millimeter.

26
00:01:36,853 --> 00:01:42,912
Next, the parts. 137 of them, and 103 get placed by their own pads.

27
00:01:42,932 --> 00:01:49,055
Markings get read where they're legible, like this TI DRV8323S.

28
00:01:49,075 --> 00:01:53,601
Most small ceramics carry no marking at all, so their values just aren't in

29
00:01:53,621 --> 00:01:56,511
the photos. The flow says so, and doesn't guess.

30
00:01:57,084 --> 00:02:00,718
Silkscreen and circuit role sort them into 11 blocks.

31
00:02:00,738 --> 00:02:06,145
Power entry, the power stage, sensing, the logic supply, the gate driver and

32
00:02:06,165 --> 00:02:11,144
the microcontroller. Then CAN, AUX, the fan, motor temperature and the

33
00:02:11,164 --> 00:02:12,215
mounting holes.

34
00:02:12,785 --> 00:02:17,374
Now the hard part. Under solder mask, copper looks a little lighter than bare

35
00:02:17,394 --> 00:02:20,008
laminate, and thin dark lines mark the gaps.

36
00:02:20,028 --> 00:02:24,316
The reader follows those gaps, and the board comes apart into islands.

37
00:02:24,336 --> 00:02:28,208
Here's the published version next to this one. Fine tracks that used to be

38
00:02:28,228 --> 00:02:29,522
missing now come through.

39
00:02:29,542 --> 00:02:34,034
That's 828 pours and 2,588 tracks.

40
00:02:34,610 --> 00:02:39,082
Looking closely found a real defect. The dark side of each dome was being cut

41
00:02:39,102 --> 00:02:40,190
out of solid pours.

42
00:02:40,210 --> 00:02:45,056
So the reader learned what a dome's shadow looks like, and now those pours

43
00:02:45,076 --> 00:02:49,331
read solid. Domes with a hole beside them dropped from 396 to 68.

44
00:02:49,916 --> 00:02:54,713
Vias get a shading test. A real dome is lit on one side and shadowed on the

45
00:02:54,733 --> 00:02:57,603
other, in the same direction as all the rest.

46
00:02:57,623 --> 00:03:02,030
The published filter ringed silkscreen and missed most of the domes. The new

47
00:03:02,050 --> 00:03:03,020
one sits on them.

48
00:03:03,040 --> 00:03:08,575
1,329 kept. The top face shows none it can trust, so these all come from the

49
00:03:08,595 --> 00:03:12,010
bottom. A few are still missed along the edges.

50
00:03:12,586 --> 00:03:17,282
It doesn't stop at copper. Each pixel gets sorted by material: mask,

51
00:03:17,302 --> 00:03:20,334
silkscreen, metal, or a part sitting on top.

52
00:03:20,354 --> 00:03:25,516
Silkscreen, mask openings, exposed metal, holes and the outline all get read

53
00:03:25,536 --> 00:03:29,471
from the photos, then written into their own KiCad layers.

54
00:03:30,049 --> 00:03:34,533
All of it goes into a real KiCad project, as photographed, before anything

55
00:03:34,553 --> 00:03:38,488
gets guessed. Here it is over the top photo, and over the bottom.

56
00:03:38,508 --> 00:03:44,060
Every stage opens as an interactive board. Pick a net and all of it lights up:

57
00:03:44,080 --> 00:03:46,346
ground, the input rail, a phase.

58
00:03:46,926 --> 00:03:51,908
Then a separate check compares what KiCad draws with what the photos show, on

59
00:03:51,928 --> 00:03:54,052
tiles the reader wasn't tuned on.

60
00:03:54,072 --> 00:03:58,933
Green agrees. Red is drawn but not seen. Blue is seen but not drawn. Purple is

61
00:03:58,953 --> 00:04:01,436
hidden under a part, so it isn't scored.

62
00:04:01,456 --> 00:04:07,147
The bottom scores 0.87. The top, 0.82, on only about half the board, because

63
00:04:07,167 --> 00:04:08,725
parts cover the rest.

64
00:04:09,286 --> 00:04:13,776
Next, every piece of copper gets a net, and a record of how it got it. Pieces

65
00:04:13,796 --> 00:04:17,290
get stitched into conductors wherever no gap separates them.

66
00:04:17,310 --> 00:04:22,152
Green means a pad touches it. Blue, it's joined through a via. Yellow is

67
00:04:22,172 --> 00:04:25,325
inferred, orange only by group. Red has no net.

68
00:04:25,345 --> 00:04:30,748
Three quarters of the copper area now has direct evidence. But 47% of track

69
00:04:30,768 --> 00:04:33,929
length is still red, so that goal isn't met.

70
00:04:33,949 --> 00:04:38,335
A schematic comes out too. All 77 nets are still marked inferred.

71
00:04:38,916 --> 00:04:43,439
Now the connections nobody can see. The flow has its own router, and it uses

72
00:04:43,459 --> 00:04:44,515
every inner layer.

73
00:04:44,535 --> 00:04:49,151
Real vias pass through pours, so crossing copper from the photos is expensive,

74
00:04:49,171 --> 00:04:52,242
not forbidden. And the length crossed gets reported.

75
00:04:52,262 --> 00:04:57,792
At 6 layers, 45 of 253 connections stay open. The white crosses are where it

76
00:04:57,812 --> 00:05:00,347
got stuck, mostly at the two chips.

77
00:05:00,923 --> 00:05:03,850
That difficulty is how the layer count gets proposed.

78
00:05:03,870 --> 00:05:11,579
2 layers cross 454 mm of seen copper. 4 layers, 384.

79
00:05:11,599 --> 00:05:16,752
6 layers, just 11. 8, about 8. Past 6, more layers stop helping. So the

80
00:05:16,772 --> 00:05:17,772
proposal is 6.

81
00:05:18,342 --> 00:05:22,814
Confidence isn't self graded. Auditors who did none of the work judge random

82
00:05:22,834 --> 00:05:25,651
crops, and some of the claims are planted false.

83
00:05:25,671 --> 00:05:30,315
Pass one of those, and the whole sample gets thrown out. That actually

84
00:05:30,335 --> 00:05:31,980
happened here, on blocks.

85
00:05:32,000 --> 00:05:37,248
Identity clears 90. Vias, blocks and placement land around 60. Values are at

86
00:05:37,268 --> 00:05:40,437
34, because most small parts carry no marking.

87
00:05:40,457 --> 00:05:45,360
Copper scores 9, because through smooth green mask, an auditor mostly can't

88
00:05:45,380 --> 00:05:50,348
tell copper from bare board. Layers stop at 30 without outside evidence. And

89
00:05:50,368 --> 00:05:54,681
connectivity is 0, because KiCad finds more missing than expected.

90
00:05:54,701 --> 00:06:00,047
Overall is the lowest category, so the loop reports stalled, at 0. It did not

91
00:06:00,067 --> 00:06:04,716
reach 90. From photos alone, it can't: values, layers and nets need

92
00:06:04,736 --> 00:06:06,876
measurements on the real board.

93
00:06:07,452 --> 00:06:11,884
Whatever is still unknown gets a test. 119 of them, each saying what to

94
00:06:11,904 --> 00:06:16,211
measure, what result to expect, and what to change if it's different.

95
00:06:16,785 --> 00:06:21,462
We did find the real design. It went in a sealed folder as an answer key. The

96
00:06:21,482 --> 00:06:25,793
flow never reads it, the agents never see it, and only totals come out.

97
00:06:25,813 --> 00:06:32,884
Real vias went from 42% to over 95%. Vias found, from 26% to 75%.

98
00:06:32,904 --> 00:06:38,935
Copper match went from 61% to 67% on top, and from 80% to 89% on the bottom.

99
00:06:38,955 --> 00:06:40,846
And 6 layers is correct.

100
00:06:40,866 --> 00:06:46,150
Pin nets barely moved. 46% to 48%. That's the weakest result here.

101
00:06:46,718 --> 00:06:52,629
So this isn't a finished board. 45 connections stay open, and KiCad counts 441

102
00:06:52,649 --> 00:06:53,238
missing.

103
00:06:53,258 --> 00:06:58,560
Finer copper also made routing harder. New design rule errors at 6 layers went

104
00:06:58,580 --> 00:07:03,540
from 6 to 594. And these thresholds came from this one board, so a second

105
00:07:03,560 --> 00:07:05,109
board has to come next.

106
00:07:05,681 --> 00:07:09,822
Every stage stays in the flow window, so you can step back through the whole

107
00:07:09,842 --> 00:07:11,848
run and open any board along the way.

108
00:07:12,442 --> 00:07:19,318
From the first prompt to this point took about 13 hours of AI time in the main

109
00:07:19,338 --> 00:07:23,827
thread, with idle gaps cut out, plus 46 sub-agents.

110
00:07:23,847 --> 00:07:28,739
About 3.4 million output tokens and 634 million cache reads, across 1,811

111
00:07:28,759 --> 00:07:33,718
model calls. Next up: real measurements on this board, and a second board.