Form-extraction pipelines
Multi-line catalog reconciliation
Line-item extraction matched against an internal catalog, with threshold validation on every row.
Read the caseAll case studies
Case study · Catalog curation
Meridian Operations · catalog with audit-friendly granularity96% accuracy. One item per turn.
The Product Specialist proposes, cites its sources, and waits for approval. It doesn't batch. It doesn't guess on ambiguous cases. Every change has a name on it.
0%
Classification accuracy on first-pass acceptance
0
Item per turn — bulk operations off by design
0x
Faster per-item curation vs. manual research
Built with
The problem
Spreadsheet bulk-edits propagate the wrong formula. Audit rows record who, not why.
Catalog quality used to be one of those things that quietly degraded when nobody owned it. A spreadsheet bulk-update would rewrite 300 rows correctly and 12 rows wrong; the wrong ones surfaced months later as variance complaints. Six months on, the curator who made a wrong code didn't remember whether it was deliberate or a copy-paste.
- The bulk mistakeCatalog updates were done in spreadsheet exports: filter, edit, re-import. A wrong formula in column G propagated to 300 rows before anyone noticed. The correction took longer than the original error.
- Classification research sprawlFinding the right code for a new product meant opening three browser tabs, cross-referencing two reference tables, and checking what similar products in the catalog used. A curator who knew where to look took 12 minutes. A curator who didn't took 40.
- Audit rows with no reasoningThe database recorded who changed a field and when, but not why. A wrong code six months old had no trail — was it deliberate? A copy-paste error? A reversed policy? The curator who made it might not remember.
- Enrichment suggestions with no contextWhen the extractor's output differed from the catalog value, a suggestion was generated. A curator looking at it had to decide: better answer, or document-reading mistake? Without context, a coin flip.
Catalog audit log · who, not why
5 unrecoverablebulk-import 312 rowsoperator A · 4mo ago12 rows wrong; reason unknown
code 8471.30 → 8471.40operator B · 6mo agono reasoning recorded
alias merged 'Cab Sav' / 'Cabernet Sauv.'operator C · 3mo agotwo distinct products
threshold 1L → 1.5Loperator A · 8mo agopolicy reversed since
enrichment suggestion ignoredoperator D · 2mo agooutcome not logged
Audit row says who · not whyNo reasoning trail
The pipeline
From inbox to verified record in one pass
Six steps per item. The one-item-per-turn limit isn't a UI choice — it's the audit model.
01Pick
Product specialist selected
Curator opens the assistant and picks the product specialist persona. Grant required to see the option.
02Intake
Item identified before action
Curator names the product, pastes an enrichment suggestion, or describes what they want to create. Specialist confirms it understood the item before proceeding.
Claude
03Research
Reference tables and neighbors pulled
Specialist pulls reference-data lookups, alias candidates, and prior classification decisions for products with similar descriptions. It does not propose yet — it gathers.
04Propose
Code + sources presented
Specialist proposes a classification code, alias additions, and metadata updates. Each proposal includes the source(s) that drove it: which reference table, which neighbor, which signal.
Claude
05Confirm
Curator approves item-by-item
Curator approves, edits, or rejects each proposal individually. No batching. Rejected proposals trigger a follow-up question; the correction logs as a new example.
06Commit
Audit row with reasoning, not just identity
Approved changes write to the catalog. Audit row includes curator identity, specialist's reasoning, source citations, timestamp. The reasoning is part of the record, not a transient log.
PostgreSQL
Similar to existing catalog items
Approve in under two minutes with full audit row
Specialist finds close matches, proposes a classification code with high confidence, and cites the matching catalog items as sources. Curator approves quickly; the audit row is complete with reasoning attached.
Novel product, no neighbors
Show candidates, ask the curator to pick
Specialist flags that it is working from reference tables only, shows the two or three candidate codes with confidence scores for each, and asks the curator to select. It does not pick one and present it as a recommendation. The curator's selection becomes the example for future similar products.
Validation review
First-pass acceptance per curation action
The specialist handles five action types. The enrichment-suggestion review is the lowest-confidence one — and the source explains exactly why.
Field-level confidence
Pass 2 — Claude self-review
HighMediumLow
New product creationFirst-time catalog entry
98%High
Classification code revisionReplace an existing code with a new one
94%High
Alias additionAdd a known-alternate description
97%High
Threshold adjustmentNumeric per-account threshold change
96%High
Enrichment-suggestion reviewAccept or reject an extractor-derived correction
64%Low
Routed to human review. The specialist must judge whether the correction was an isolated document-reading error or a systemic catalog gap. Adding a per-product-type accuracy signal to the tool call would raise this substantially — it's a data-availability problem, not a model problem.
4 of 5 fields cleared the 0.85 threshold
model: First-pass acceptanceThe stack
Boring tech, glued together well
Each vendor handles what it's best at. Aisyst owns the orchestration layer in between.
Claude
Plans research, proposes codes, and writes the reasoning summary on the audit row
PostgreSQL
Catalog records, alias table, audit rows with reasoning, enrichment suggestions
Drizzle ORM
Type-safe queries against the catalog and reference tables
Microsoft Teams
Notification surface when low-confidence enrichment suggestions accumulate
Third-party logos are trademarks of their respective owners and appear here only to indicate integration.
Outcomes
What changed when 'who' became 'who and why'
Catalog reasoning lives on the audit row, not in someone's memory. A wrong code six months from now is investigable.
0%
Classification accuracy on first-pass acceptance
0
Item per turn — bulk off
0x
Faster per-item curation
0%
Catalog changes audit-trailed with reasoning
Watch first-pass acceptance per proposal
Currently 96%. If curators are editing more than 4% of proposals, two failure modes are possible: the reference data is stale (fix with a refresh) or the specialist's prompt is missing examples for a new product category (fix with a few new examples). Both are easier to address at 4% than after the catalog has accumulated 200 wrong codes.