Given the characteristic list, AutoBalloon resolves tolerances, numbers characteristics clockwise-from-12, and emits the ballooned drawing PDF plus the AS9102 FAIR workbook with no manual steps.[F001][F002][F003][F004] Demonstrated end-to-end on a real aerospace FAI package.[F005]
Scope today: one part, eight characteristics, vector PDF only. Three sheets. Sheet classification keys on text-layer density.[F006]
An AS9102 First Article Inspection Report starts with ballooning: reading the drawing, marking each dimension, note and tolerance with a numbered balloon, then re-typing all of it into Form 3 in the same order.[F007] It is slow, it is repetitive, and a transcription slip in either direction is an error nobody catches until the customer does.
Every inspectable characteristic on the drawing — dimensions, tolerances, GD&T frames, and the notes on the parts list — has to be found and marked. Nothing on the sheet says which entries are inspectable; a QA engineer decides.[F007]
Balloon numbers are not arbitrary. They have to run in a stable order so a second person can walk the drawing and find characteristic 6 without hunting — clockwise about the view, with the parts-list notes taken first.[F001]
Each numbered characteristic becomes one row of Form 3, carrying its reference location, designator and requirement. The same values are keyed a second time, by hand, in the same order.[F007]
That distinction runs through this entire page, and it is the reason AutoBalloon is a narrow tool rather than a broad one. A drawing can tell you what a characteristic is supposed to be. It can never tell you what the part actually measured.[F008]
If you have never filled one in: an AS9102 First Article Inspection Report is not a single document. It is three forms, each answering a different accountability question, and each fed from a different place. Form 1 accounts for the part, Form 2 for the product, Form 3 for every characteristic — one row each.[F007]
Solid lines carry data that exists somewhere in the drawing package. Dashed lines carry data that does not — it arrives from the shop floor, or from a person signing.[F008]
Forms 1 and 2 are broadly a header exercise: a fixed set of fields, filled once per part. Form 3 grows with the drawing. A print with two hundred inspectable characteristics is two hundred rows, each one found by eye, numbered by hand, and keyed twice.[F007]
AutoBalloon writes all three, with Form 3 rows generated from the characteristic list rather than typed.[F004]
This is the sharpest thing we learned building AutoBalloon, and it decides what a tool like this can honestly do. The FAIR task splits cleanly into three data classes. One is extractable from the drawing package. One never appears on a drawing at all, at any point, for any part. One is a human putting their name to something.[F008]
source: the PDF package
This class is the ballooning engine's whole job. Everything AutoBalloon claims to do lives inside this column.[F008]
source: the shop floor — never the drawing
No drawing has ever contained any of this. It cannot be extracted, inferred, or generated — it has to be measured and merged in.[F008]
source: a person
Always manual, by design. A signature that a program can produce is not a signature.[F008]
When AutoBalloon writes the AS9102 workbook, the inspection-result columns of Form 3 are left empty on purpose.[F004] Read against the three classes above, that is not a missing feature — those columns are Class 2, and Class 2 values do not exist anywhere in the input. A tool that filled them from a drawing would be inventing them.
Merging real CMM results into those columns is a separate job, with separate inputs. It is not built.[F004]
We assumed a vector PDF drawing would carry a text layer we could read dimensions out of. Ingesting the gold package showed otherwise: sheet 0 — the drawing itself — yielded zero text tokens. The CAD system had exported every dimension, every title-block entry and every existing balloon as vector outlines. There is no text to read; there are only strokes that look like text.[F009]
Token counts are a property of this drawing package as it was exported, not a measure of anything AutoBalloon does.[F009]
The obvious reading of a sheet with almost no text is "this is a cover page, skip it". The correct reading is the opposite: a near-empty text layer is precisely the marker that says this sheet carries the drawing and needs vision. Sheet classification is built on that density signal.[F009][F006]
A vision-only engine was ruled out on architecture rather than preference. Drawing a leader line requires an exact (x, y) coordinate on the page; vision models return approximate boxes. That mismatch is structural, so the design is hybrid — semantic reading for robustness to GD&T and messy layout, deterministic PDF parsing for coordinates.[F011]
Because the drawing sheet has no text layer, balloon coordinates on that sheet cannot come from token matching — they have to come from vision boxes or vector-geometry analysis.[F010] The vision path exists in the code but has never been executed or graded.[N003] That is the single largest gap between what is designed and what runs.
Most dimensions on a real print carry no tolerance next to them. The tolerance is implied, and lives in a table in the title block, indexed by how many decimal places the dimension is written to. Resolving that is deterministic text work, and it is the piece that later turns a measured value into ACCEPT or REJECT.[F013]
Deterministic text parsing with independent unit tests — this is one of the few places on the page where we are confident, and it is confidence about a unit-tested rule, not about performance on your drawings.[F002][F013]
Symmetric ± callouts such as Ø 27.0 ± 0.1 resolve directly from the callout text. Bare dimensions resolve through the title-block table as shown above. Both are deterministic text parsing covered by independent unit tests.[F002][F013]
Explicit min/max pairs and asymmetric +x/−y callouts have no parsing code and no tests at all. The word "unilateral" appears once in the source, in a doc comment describing an aspiration. If your prints use those forms, this will not resolve them.[N005]
Each of these is implemented and covered by unit tests. Everything outside this list is on the ladder further down, marked as not built.
Characteristics are numbered clockwise from 12:00 about the view centroid, with PART LIST notes ordered first. The sort is unit-tested over synthetic coordinates unrelated to any real drawing.[F001]
Symmetric ± callouts resolve directly. Bare dimensions resolve against the title-block general-tolerance table by decimal place — 2.7 becomes ±0.3. Deterministic text parsing with independent unit tests.[F002]
Numbered balloons and leader lines are drawn onto the source drawing and written out as a PDF you can hand to an inspector.[F003]
Form 1 part accountability, Form 2 material and process from the BOM, Form 3 characteristic rows. Inspection-result columns are left blank on purpose — those values never appear on a drawing.[F004]
The pipeline ran end-to-end once, on a real aerospace FAI package — 72257339-001 "RESTRICTOR, OVERFLOW", program ST AERO - 722, Hampson Industries — and produced output matching the human-produced Form 3.[F005] Read that as a demonstration, not a measurement: the input for that run was hand-authored with the answer key in context, so it cannot stand as evidence about detection.[N001]
These stages run in order, so they are numbered. Stage 02 carries its caveat on the node itself rather than in a footnote, because the caveat is what makes the rest of the sentence true.
A vector or text PDF is read and its sheets classified by text-layer density. Vector PDFs only, three sheets, one part.[F006]
The characteristic list — every dimension, note and tolerance to be inspected — enters the pipeline here.
Today this list is produced by a person, not by the software. The graded run performs no detection: it reads a hand-authored list from disk.[N001] A vision path exists in the code but has never been executed or graded.[N003] There is no operator confirm-and-edit workflow.[N004]
Each characteristic is tied to a point on the sheet, and its tolerance resolved — symmetric ± directly, bare dimensions against the title-block table.[F002]
Characteristics are ordered clockwise from 12:00 about the view centroid, PART LIST notes first, and assigned their balloon numbers.[F001]
Balloons and leader lines are drawn onto the drawing[F003] and the AS9102 workbook is written — Forms 1, 2 and 3, inspection results left blank.[F004]
AutoBalloon was developed against one real, complete aerospace FAI package: part 72257339-001, "RESTRICTOR, OVERFLOW", program ST AERO - 722, Hampson Industries. Because that package arrived already ballooned and with its forms filled in by a human, it works as an answer key. Here is the whole of it — not a selected extract.[F012]
| № | Section | Where it lives | Requirement | Note |
|---|---|---|---|---|
| 1 | Part list | BOM sheet | 1N | — |
| 2 | Part list | BOM sheet | 2N | 722-STA-PS-125 SEAL B |
| 3 | Part list | BOM sheet | 3N | 722-STA-PS-163 |
| 4 | Part list | BOM sheet | 4N | 722-STA-PS-194 |
| 5 | Measure | Drawing sheet | 2.7 ± 0.3 | Bare dimension — tolerance from the title-block table[F013] |
| 6 | Measure | Drawing sheet | Ø 27.0 ± 0.1 | Symmetric callout[F002] |
| 7 | Measure | Drawing sheet | Ø 44.5 ± 0.3 | Symmetric callout[F002] |
| 8 | Measure | BOM sheet | 0.0074 KG | Weight — a measure that is not on the drawing view |
Form 3 groups characteristics: balloons 1–4 are PART LIST notes carried on the BOM sheets, and 5–8 are MEASURES. That grouping is why the numbering rule puts parts-list notes first and only then walks the drawing clockwise from 12:00.[F012][F001]
Characteristic 5 is a bare 2.7 whose limits come from the title block; 6 and 7 carry their own symmetric ± callouts. One small part exercises both resolution paths, which is why the tolerance module was the first thing written and unit-tested.[F013][F002]
This is the entire evidence base: one part, eight characteristics, vector PDF only, three sheets.[F006] The pipeline ran end-to-end on it once and produced output matching the human-produced Form 3.[F005]
Treat that as a demonstration and not a measurement. The input for that run was hand-authored with the answer key in context, and the source drawing already carried balloons 1–8 — so both the detection and the numbering-order evidence from this part are circular.[N001] Header extraction on this part also matches literal part-specific strings and will not transfer to another drawing without code changes.[N006]
Ballooning is a solved problem at the top of the market. HighQA, Net Inspector and iQA all automate it, and they are quoted at roughly ₹7–8 lakh per licence plus annual maintenance.[M001] For a supplier running a handful of first articles a quarter, that licence costs more than the work it removes — so the work stays manual.
Market pricing for the established commercial FAI and ballooning tools.[M001] That is a competitor price, gathered during market research. It is not a performance figure and it is not a measurement of anything AutoBalloon does.[M001]
Source note: the ₹7–8 lakh figure describes what other vendors charge, nothing more. No claim about AutoBalloon's capability, quality or output is derived from it, and nothing on this page compares AutoBalloon's results to theirs — we have no basis for such a comparison and do not make one.[M001]
AutoBalloon is not a competitor to those suites today. It does four things deterministically on vector PDFs for one part shape, and the ladder below lists everything it does not do.[F006]
The parts of ballooning that are pure transcription — resolving a tolerance, ordering the balloons, writing Form 3 rows — are deterministic and testable.[F001][F002][F004] Those are the parts built first.
Deciding which entries on a print are inspectable remains a person's job here. There is no operator confirm-and-edit workflow, and the autonomous path has never been run or graded.[N004][N003]
Most tools in this category describe the top rung and let you assume the bottom two. Here they are labelled, so you can tell which one you would be buying into.
Implemented, deterministic, and covered by unit tests. These are the only claims on this page that describe working software.
SCOPE — one part · eight characteristics · vector PDF only · three sheets.[F006] One end-to-end demonstration run on a real FAI package.[F005]
None of the following exists as working software. Some has no code at all; some has code that has never been run or graded. Either way, do not plan around it.
TEST DEBT — no end-to-end test exists; the only in-code scoring assertion grades the gold sample against itself.[N002]
Direction of travel only. Nothing on this rung is built, scheduled, or promised — it is here so you know what we are aiming at, not what you would receive.
ASPIRATIONAL — no code, no dates. When any of this lands it moves up a rung, and the fact sheet changes first.
AutoBalloon is early. The useful next step is a real drawing from a real first article, so we can find out what breaks outside the one part it has seen. Tell us what you inspect and we will tell you honestly whether it is in scope today.