Find the substitute before you buy the duplicate

In large inventories the same kind of part often sits under different names, units and spec formats. Buyers can’t see the equivalent already on the shelf, so they order new stock.

An AI pipeline that cleans messy inventory master data, merges duplicates and finds functional substitutes for the item you need. Parameters first; a category-expert LLM steps in only where the algorithm is unsure.

Talk about your catalog

SchematicLLM Analog Finder

Production numbers from an industrial company. Client name under NDA.

76kcatalog items run through the pipeline at an industrial company

1 in 3items got a substitute or a conditional match

72%of “analogs” in electrical were duplicates of the same part

2 daysto onboard a new product group; the first build took 3 weeks

Built: My idea and architecture. I wrote the first prototype, for cables, from scratch in Python and JavaScript
Scaled: In production at an industrial company. A team of engineers and AI agents scaled it under my lead
Tested: Open retail catalogs; synthetic inventory data in SAP and other ERP formats

What it proves for your business

One architect with AI agents builds a working prototype; a team scales it on a proven design. A new product group becomes configuration, not development.

Schematic 04-AFrom a messy catalog to a justified substitute

Stage I · Data first

01 Normalize Raw records cleaned: units, abbreviations, spelling.
02 Parameter template For each item class, the parameters that matter, set by experts’ prompts.
03 Enrich Every parameter filled from the web.
1. aManufacturer catalogs
2. bDistributor price lists
3. cGrounded web search
next source only if the previous one is not enough
04 Standardize Records rewritten on the filled parameters.
05 Deduplicate The same item under different names becomes one record.

Stage II · Algorithm first, LLM only where in doubt

06 Parametric matching An algorithm looks for analogs of every item by its parameters.

Clear match

In doubt

07 Category-expert LLM A separate expert prompt per category.

Electrical
Bearings
Cables
Fasteners
…

08 Verdict Analog or not · similarity, % · a written comment comparing the parameters of the original and each candidate.

Schematic 04-BIn production · industrial company · client under NDA

1 in 3 items got a substitute Share of the whole catalog, after one pass of the pipeline.
- Strict analog · 8%
- Conditional, with what to check · 25%
- No substitute · 67%
72% of “analogs” were duplicates Matches in electrical: the same part number filed under different codes. Found by a rule, no model needed.
- Duplicate · 72%
- True substitute · 28%
The AI judge: open model beats paid ones Agreement with a reference set of 240 pairs.
- Open model inside the perimeter 82–84%
- Paid frontier models 73–79%
0–1 dangerous errors per 120 pairs. Paid models missed about 30% of real analogs. The judge runs inside the company perimeter.
Where to search beats which model searches Missing parameters filled on the same 99 items.
- Targeted search + in-house model 41%
- Paid AI search 11%
3.6× more found, about 50× cheaper.

173kproducts from 4 manufacturer catalogs, at no cost
665kmodel calls inside the company perimeter
1,683matching rules generated from expert prompts

The problem

In large inventories the same kind of part often sits under different names, units and spec formats. Buyers can’t see the equivalent already on the shelf, so they order new stock. Finding a substitute for a scarce item is manual work that depends on the memory of the most experienced specialist.

In production

The approach runs at an industrial company on its own master data. The client name is under NDA.

76k catalog items went through the pipeline. 1 in 3 got a strict analog or a conditional substitute with a note on what to check.
72% of matches in electrical were duplicates: the same manufacturer part number filed under different codes. A rule finds them without a model, and they can be merged right away.
1,683 matching rules across 301 categories, generated from expert prompts: what must match, what must be no worse, what is only taken into account.
815k parameter values and 2.1 million specs extracted from web pages; 220k candidate pairs judged.
The judge is an open model inside the company perimeter: 82–84% agreement with the reference set, 0–1 dangerous errors per 120 pairs. Paid frontier models scored 73–79% on the same test and missed about 30% of real analogs. Judging, the bulk of the work, runs inside the company.
Paid LLMs for the whole project cost less than the first estimate for web search in a single product group. Everything else ran as 665k model calls inside the perimeter.
Where to search beats which model searches. Targeted search filled 3.6× more missing parameters than a paid AI search, at about 1/50 of the cost.
Full manufacturer catalogs instead of item-by-item search: 173k products and 4.65 million values from four manufacturers, at no cost.
Experts teach the pipeline. Two experts labeled 559 bearing pairs; false analogs fell from 43 to 18, and no valid substitute was lost.
2 days to onboard a new product group. The first three groups took 3 weeks, while the pipeline was being built.

My role

I came up with the idea, wrote the first prototype for cables from scratch in Python and JavaScript, and designed the architecture the production service runs on. A team of engineers and AI agents scaled it to new product groups under my lead. I was the architect, technical supervisor and product manager.

That is how I work with clients: I design the approach and prove it on your data, then a team of people and AI agents scales it.

How it works

Data first. The pipeline cleans the catalog before it looks for anything:

normalizes the raw records;
builds a parameter template for each item class from experts’ prompts;
fills every parameter from the web: manufacturer catalogs first, then distributor price lists, then grounded web search;
standardizes the records on those parameters;
merges duplicates.

Then it looks for analogs. An algorithm matches every item by its parameters. Only where it is unsure does a large language model step in, with a separate expert prompt for each category: electrical equipment, bearings, cables, fasteners and so on. The verdict: analog or not, a similarity percentage and a written comment that compares the parameters of the original and each candidate.

Why it is built this way

Cost under control. The expensive model works only on the cases the algorithm can’t settle.
Explainable. Every verdict comes with the parameters behind it, so a buyer can check it.
Expertise captured. What a category expert knows lives in a prompt, not only in one specialist’s memory.

Where it applies

Manufacturing, retail, marketplaces and distribution: anywhere with large inventories and complex procurement. Procurement is only the example. The same approach works for master data in sales, finance, logistics and legal.

Try it on your catalog

Under NDA. For a first check, an anonymized export of one full product group is enough, from 1,000 items.

Want the same approach on your problem?

Email Alex All projects