AI News HubLIVE
In-site rewrite3 min read

Crossary – AI-assisted field mapping that outputs signed Excel files

Crossary is an AI-powered field mapping tool for integration engineers, consultants, and data professionals. It uses a five-stage pipeline to extract fields from source and target specs, propose mappings with evidence, and export signed Excel workbooks. It emphasizes honesty, determinism, and data privacy.

SourceHacker News AIAuthor: migueljpalmeida

Who it's for

If you hand-build the mapping sheet someone else implements — this is for you.

01

Integration & interface engineers

Mapping heterogeneous specs — PDF, XML, EDI-style guides, Excel — to a target schema.

02

Implementation & onboarding consultants

Turning a client's messy spec into a reviewed workbook on day one — not week three.

03

Data & migration engineers

Building the field inventory and design before any records move.

04

EDI & interface analysts

Turning implementation-guide PDFs into evidence-backed rows where the “why” has to be defensible later.

How it works

A strict five-stage pipeline. Then a round-trip.

Upload to signed workbook in five steps — and back again, without losing a note.

01

Artifacts

Drop in your source + target specs. If a file can't be fully read, it tells you exactly how much it dropped.

xlsx · pdf · csv · json · xml · xsd · sql · yaml

02

Fields

A field inventory is extracted from both sides — every path, every column — so you map against the real surface.

03

Mapping

For each target: a proposed source, type, verbatim evidence, reasoning, and confidence. Abstains when unsure.

04

Validation

A deterministic check for structural & cardinality blockers. Zero AI, zero spend.

05

Export

A signed .xlsx your team can open anywhere — that reconciles cleanly when it comes back.

Anatomy of a reviewed row

Six things every row tells you. Nothing taken on faith.

Open any proposed mapping and you see exactly why the AI suggested it — and how sure it says it is. The evidence, not the confidence, is the product.

target field ①

order_date *

confidence ④

high

source field ②

S_SHIP_DATE

mapping type ③

transformation

to_date(S_SHIP_DATE)

evidence — verbatim from the source ⑤

— ship date, ISO-8601 string. Target is a date type.

See it actually happen

The honest gap it won't guess — and the round-trip that holds.

The real sample, end to end: where the AI withdraws a tempting guess and asks instead, and where your edits survive a trip through Excel and back.

the sample

hubspot → salesforce

target fieldlow confidenceno confidence

Region

←Countryno clear source

No HubSpot field tracks sales territory, and it must not be guessed from Country.

?Region (sales territory) isn't in the HubSpot export. Assign it in Salesforce, or point me at a source?

When it isn't sure, it says so.

confidence: none · no guess

Mapping memory

Every mapping you approve makes the next one faster.

Sign off an export and your approved field-pairs become a private, workspace-scoped library. On the next run it gap-fills only the targets the AI abstained on — inserted as suggestions you still review, never auto-applied.

Scoped to your workspace — it never crosses to another.

Captures the decision only — no evidence, values, or client data.

It never trains a shared model.

next run · a gap, pre-filled

Industry

library

Trust & data

Trust isn't asked for here — it's enforced in code, and written down.

Guaranteed in code

Honest about what it read

If more than ~10% of a source is dropped during ingestion, it stays flagged — in the app and on the export cover sheet. A partial read is never reported as complete.

“Validated” means one thing

Validation is deterministic and checks structure and cardinality only. It does not claim semantic correctness — and the UI says exactly that.

Never loses your note

On re-import it applies what matched, skips what changed underneath you, and turns every reviewer note into a tracked question. Nothing is silently overwritten.

Schema-validated, or rejected

Every AI response is checked against a strict schema before it can touch your data. A malformed result is thrown away, never persisted.

Your data

Specs in, not your records

Crossary maps from your source and target specs — schemas, dictionaries, guides. It's built to work from the spec, not your data, so in most cases there's no need to upload production records or PII.

Pricing

Start free. Pay for AI runs, never for reviewing.

Reviewing, validating, exporting, and round-trip re-import don't call the AI — so they're always free, on every plan.

Free

$0

For your first real integration.

Start free

3 integration credits

Free review, validate & export

Round-trip re-import

Mapping Memory included

1 workspace · up to 2 members

Most teams start here

Pro

$99/ month

~20 integrations a month.

Start with Pro

Everything in Free

~20 integration credits / mo

Higher-accuracy pass

Unused credits roll over

3 workspaces · up to 3 members

Team

$399/ month

~75 integrations a month.

Start with Team

Everything in Pro

~75 integration credits / mo

Unlimited workspaces · up to 10 members

Shared mapping library

Plus applicable taxes.

Run out? Existing work is never blocked — review, validate, export & re-import stay free. Unused credits roll over, you can buy more anytime, or upgrade. Only new AI runs pause.

Questions, answered straight

The honest answers, before you sign up.

What's a credit?

Roughly one standard integration. Higher Accuracy and unusually large or heavily re-run jobs use more — and you can see what each integration cost.

What's always free?

Reviewing, validating, exporting, and round-trip re-import. They don't call the AI, so they never use a credit.

Is accuracy gated to higher plans?

Every plan gets the same core mapping quality. Higher Accuracy — an optional, heavier pass — is on paid plans; Mapping Memory is included on every plan.

What happens when I run out?

Existing work is never blocked — review, validate, export and re-import stay free. Unused credits roll over, you can buy more anytime, or upgrade. Only new AI runs pause.

How do I upgrade?

Click subscribe on the plan you want — secure checkout, and your plan activates immediately.

Does my data train your model?

No. Mapping Memory is scoped to your workspace and never crosses it; no shared model learns from your mappings.

Map your first integration in the next ten minutes.

Upload a source and a target spec. Get back a reviewed, evidence-backed workbook — with every honest gap flagged for you.

Start free

3 integration credits free · no card · no lock-in.