01
Read
Takes the export as it comes. Columns are mapped automatically, and the original layer is stored exactly as uploaded — it is never modified. Every change lives in a separate cleansed layer, beside the original.
01.a.1SEI · Flagship
MIDAS.
Material Intelligence & Data Automation System
Your material master already knows what you own. MIDAS makes it say so — clearly, once, and with the evidence attached.
01The problem
Intelligence, trapped in text.
Every plant keeps a material master: the record of every motor, bearing, valve, seal, spare and consumable it buys, stores and maintains. Built up over decades by many hands, it fills with the same part written five different ways, specifications buried in free text, missing manufacturers and mixed units.
MIDAS reads those records the way an experienced materials engineer would. It structures them, finds the duplicates text search can't see, reasons about which parts can genuinely stand in for which, and puts a gate in front of new codes so the mess stops growing.
- The same part entered many times, each under a different description
- Specifications locked inside free text no system can filter on
- Missing or inconsistent manufacturers, units and classification codes
- Parts filed as the machine they belong to — a brush holder filed as a motor
- No quick, trustworthy answer to “do we already have this?”
- New codes raised every day, so every clean-up starts decaying at once
What it does, in one record
Raw text, as found
BRG TPR RLLR DBL ROW ID 510MM OD 950MM MAKE: SKF AG
Structured, with its source
- class
- bearing
- item role
- is a bearing — not a part for one
- bearing type
- taper roller
- rows
- double
- inside diameter
- 510 mm
- outside diameter
- 950 mm
- manufacturer
- SKF
- source
- rules engine · pattern + decoder
Illustrative example. The original text is kept unchanged beside the cleansed record.
02How it works
Seven passes.
Rules first, reference knowledge next, language models last — and only when the first two can't settle a record.
02
Harmonise
One quantity system for every comparison. Horsepower, PSI, Fahrenheit and kilovolts reduce to canonical units before any match is made, and every conversion is logged with its original value. Designation scales such as pressure class are compared on their own scale, never converted into numbers they don't mean.
03
Classify
A weighted, multi-signal classifier assigns the material class — and separately reads the item's role. SAP text is head-noun-first, so MIDAS can tell a thing that is a motor from a thing built for a motor. When it can't tell, it says “generic” rather than guess.
04
Extract
Pulls structured specifications from the free text using engineering patterns and manufacturer type-code decoders, with a schema per material class. Every value carries its source and a confidence.
05
Fill gaps
Missing specifications are filled from a curated reference catalogue first, never from the open internet. Inference needs enough known specs and enough confidence; below that the field stays blank and is flagged. A visible blank is recoverable. A confident wrong value is not.
06
Deduplicate
Exact duplicates are found by a fingerprint of the cleaned specifications, so records that look different but are the same part collapse together. Functional duplicates are judged both ways, part against part, before they are ever grouped.
07
Review
Anything uncertain goes to a review queue with its evidence and a timer. MIDAS changes no record until a person approves it.
Two layers, side by side
Every upload becomes two versions: the original, exactly as it arrived, and the cleansed record MIDAS builds from it. Duplicates are matched on the cleansed layer — so two records that looked different on the day they were typed can turn out to be the same part.
03Reasons like an engineer
Can this stand in for that?
Sharing specs isn't the same as being interchangeable. A substitute is valid only when it meets or exceeds every rating and fits every interface exactly — and raising a rating often changes an interface.
R01
Capability can exceed
A seal rated to 320 °C can serve where 280 °C is needed. Insulation class H can stand in for F. Ratings are directional, and MIDAS knows which way each one points.
R02
Interfaces must match
A higher pressure-class flange is stronger and still won't bolt into the same line. Bore, flange, mounting, voltage and phase match exactly — capability rose, fit was lost.
R03
Bigger isn't free
An oversized motor isn't an upgrade. Its overload relay, contactor, cable and breaker were all sized to the original nameplate. A rating other equipment was sized against never upgrades for free.
R04
Kind before spec
An actuator is not a motor, however much the specs overlap. Equipment-kind walls mean different equipment never merges or substitutes, while the same equipment written two ways is recognised as one.
R05
Traps, encoded
IP67 doesn't imply IP65 — they're different tests. A higher-efficiency motor can draw far more starting current and weld an undersized contactor. These go out as conditions, never silent passes.
R06
Built-for is reported
Parts written “for” a specific machine are kept apart from general-purpose stock and labelled with what they were built for — shown, never hidden, and ranked below equal general parts.
Verdicts, not scores
Identical
The same part under another code.
Substitutable
Meets or exceeds every rating, and every interface fits.
Conditional
Could serve — with the condition stated on the row.
Not substitutable
Refused, with the reason named.
Undetermined
Not enough data to say. MIDAS says so.
Every verdict names the rule it used, so a reviewer can overrule it.
04What you get
The whole lifecycle.
From the first clean-up to every new code raised after it.
01
Search-first creation
A new material code has to be checked, justified and reviewed before it exists.
- Candidates found by specification, not wording — the same motor written five ways still turns up
- Submission is gated when an existing part already matches on its critical specs
- The only way past the gate is a written justification, with reason codes
- The evidence the requester saw is frozen, so the reviewer sees exactly the same thing
- Alternatives that differ but would serve are offered too, each with its verdict
02
Smart search
Ask by part number, by description or by specification, and get a decision.
- Hybrid of semantic similarity and tolerance-aware attribute matching
- Clear outcomes: reuse what exists, review a close match, or create new
- Exact matches and genuine alternatives shown in separate sections
- Every result explains why it ranked where it did
03
Procurement gate
Checks the master and the stock before a purchase order is raised.
- Proceed, review, redirect to an existing code, or block
- Looks for in-stock alternatives before buying again
- Every verdict carries its reasoning
04
Copilot
Ask questions of your material data in plain language.
- Plans its answer from MIDAS's own tested tools
- Strictly read-only by construction — it has no way to write
- Shows the data behind every answer
05
Standards & SAP
Clean descriptions and codes, ready to load back.
- Standardised short and long descriptions
- UNSPSC mapping from the authoritative reference — unverified codes stay unassigned
- SAP-aligned material types, groups and units
- Three exports: human-reviewed only, human plus high-confidence automatic, or everything
06
Related parts
Knows how classes connect — what drives, protects and fits what.
- A typed engineering graph between material classes
- Join checks such as coupling torque and motor current, always labelled when estimated
- Associations it can't verify are shown as such, never upgraded
05Gets sharper with use
Learns from people, not around them.
- L1
A correction is memory
Once a person corrects a value, MIDAS keeps it through every future reprocess. It doesn't repeat tomorrow the mistake it was corrected on today.
- L2
Every decision is attributed
Approvals, rejections and justifications go into one ledger: who decided, what, and why.
- L3
Patterns become proposals
When the same judgement is confirmed independently enough times, it becomes a proposal for a person to accept. Nothing is ever applied on its own.
- L4
Measured, not claimed
Classification and substitution are scored against hand-labelled test sets, by precision and recall per class. The number that matters most is “confidently wrong”.
Sovereign by design
- Runs on your own infrastructure
- Local language models by default — zero external calls
- Works fully offline in rules mode, with no AI service at all
- External AI is an explicit, per-session opt-in
- Every value traceable to its source and confidence
- The original data is never modified
Questions about MIDAS
Does MIDAS change our data?
Not on its own. Your original upload is kept exactly as it came in, and every transformation lives in a separate cleansed layer with its source and confidence. MIDAS proposes changes; people approve them.
Do we need to send data to an AI provider?
No. MIDAS runs on your infrastructure with local language models by default, and most of its work is done by deterministic rules and a reference catalogue that call no model at all. An external AI service is something you switch on deliberately, if you ever want it.
What if MIDAS isn't sure?
It says so. A field it can't fill with enough evidence stays blank and is flagged; a class it can't decide is left as generic; a substitution it can't validate is marked undetermined. In master data, restraint is the intelligence.
Which material classes does it cover?
It started with rotating equipment — motors, bearings, gearboxes, pumps, valves and seals — and its schemas now span a broad industrial taxonomy, from switchgear and fans to couplings, transformers and wear parts. Each class carries specifications tagged as interface, rating or duty point, so the reasoning works the same way across all of them.
Does it work with SAP?
Yes. It reads standard material master exports and writes back SAP-aligned descriptions, groups, units and classification codes, in exports you choose by review level.
How do we start?
With a sample of your material master. We'll show you what MIDAS finds in it — duplicates, gaps and misfiled parts — before you commit to anything.
Bring a sample of your material master.
We'll show you what MIDAS finds in it — duplicates, gaps and misfiled parts — before you commit to anything.
Contact
Hire the
swarm.
Tell us about your material master. We'll tell you honestly what MIDAS would do with it — and what it would take.
hello@quantumbeetle.ai