Domain models, not prompts
Deterministic engines set every number. The language model drafts, explains and cites — never does the math.
CP-SAT schedulers
wafer-map vision
intermittent-demand stats
Why Semi³ · A product of Owllys
Three plus three plus three is nine.
Three plus three plus three is nine.
Three cubed is twenty-seven.
A yield answer that ignores the tool is incomplete. A supply answer that ignores the lot is fiction. Semi³ multiplies all three — which is why the number at the end is not a sum of its parts.
Semiconductor operators + AI engineers — one team
Fab operations
Equipment lifecycle
OSAT & test
Supply chain
12–48business hours to your roadmap
38products scored against your answers
5steps — instant preview in your browser
The name
Why it is cubed, not tripled
The exponent is the argument. Three factors, multiplied — and multiplication has a property addition does not.
Semi3 = Lot× Tool× Network
- Lot Wafer, die, unit, serial. Every number traceable to the silicon it came from.
- Tool Chamber, recipe, qual state, drift. Every cause traceable to the machine that made it.
- Network Supplier, OSAT, customer, contract. Every consequence traceable to the promise it breaks.
The loss is in the seams, not the steps
Nobody in your fab is bad at their job. Yield engineering reads a wafer map better than any vendor could. Equipment engineering knows every chamber by its sound. Planning protects the die bank. The money leaks between them.
A parametric shift surfaces on a wafer map on Tuesday. The chamber that drifted caused it three weeks earlier. The ship-and-debit it breaks is found six weeks later. Three teams, three systems, three versions of the truth — and nobody holding the thread that runs through all of them.
Adding tools is the wrong operation
The instinct is to buy a yield tool, an equipment tool and a planning tool. Three good products, side by side. Three plus three plus three.
The value was never in the three factors. It was in the combinations — and addition throws all of them away.
On one spine, “why did yield drop” and “what is it costing us” stop being two questions for two teams. The genealogy survives the saw: wafer to die to unit to serial to shipment to the return that comes back two years later.
And any factor at zero takes the rest with it
This is the honest test to put to every vendor in your pipeline, including us. Multiplication is unforgiving — one zero and the product is zero.
- The horizontal AI platform Lot = 0. It can do anything, so it knows nothing about yours.
- The BI dashboard Tool = 0. It reports the shift beautifully and cannot tell you which chamber caused it.
- The chatbot on your MES Network = 0. It answers about the lot and has no idea whose contract just broke.
Deep tech is what it takes to keep all three non-zero: solvers, vision models, physics-aware twins and a genealogy spine — rehearsed on a synthetic fab before they ever meet a row of your data.
Your competitors are adding tools. Multiplication is a different business.
Origin
One team. Two disciplines. Zero translation loss.
Most enterprise AI interviews the experts. Ours was built by them — one room, one product.
Semiconductor operators
Commissioned tools. Ran OSAT floors. Chased allocation through shortages. Closed automotive 8Ds.
- Fab · equipment · OSAT · supply-chain operating experience
- Workflows written by the people who ran them
- Playbooks and failure patterns encoded as working software
Advanced AI engineers
Enterprise AI platforms already in production across four industries.
- Deterministic engines set the numbers — the LLM never does
- Cite-or-abstain answers, enforced by evaluation gates
- Every agent action carries evidence and an audit trail
The vocabulary test
Click any term — the platform already speaks fab. If your AI vendor thinks MSL is a soccer league, we should talk.
What it is
A priority lot that jumps the queue — and ripples through every other commitment on the floor.
What our AI does with it
Inserts it with ripple-cost and displaced-lot analysis, honoring due dates and tool capability — recommended to your MES, never forced.
What it is
Moisture-sensitivity floor time: once unsealed, a part has a fixed window before it must re-bake.
What our AI does with it
The scheduler honors MSL floor-time clocks per lot; the die-bank optimizer tracks moisture clocks so no lot quietly expires on a shelf.
What it is
The eight-discipline corrective-action report your automotive customers demand after an escape.
What our AI does with it
Drafts the 8D from the investigation's evidence trail — cite-or-abstain — and releases it only through a named approval gate. Authoring drops from days to hours.
What it is
A supplier's product/process change notice. Buyers absorb thousands a year — each one hand-traced today.
What our AI does with it
Traces blast radius across BOMs, quals, WIP and customers in minutes, and quantifies the last-time-buy before the 90–180-day window closes.
What it is
The export-control classification that decides whether tomorrow's shipment is legal — under rules that flipped repeatedly within the last year.
What our AI does with it
Determines ECCN/HTS retrieval-first, cites the governing text, abstains when ambiguous — and re-screens your whole order book the same day a rule changes.
What it is
The buffer of known-good die held between fab and assembly — working capital parked against uncertainty.
What our AI does with it
Solves wafer bank vs die bank vs finished goods stochastically — where to hold the buffer, and in what form — with intermittent demand forecast honestly.
What it is
The spatial fingerprint of a wafer's failures — edge ring, scratch, donut — each pattern pointing to a different cause.
What our AI does with it
A vision model names every signature before a human opens the lot, feeding ranked, evidence-cited root-cause hypotheses.
What it is
JEDEC/AEC-Q supplier qualification: 6–18 months of sample lots, documents and gates before a second source is real.
What our AI does with it
Runs the qual plan as a durable project — document chase, sample-lot tracking, readiness scoring — targeting 30–50% shorter cycles.
Domain-native
What “built for semiconductors” actually means
Three design decisions — the difference between domain-native and a semiconductor landing page.
Fab vocabulary in the data model
Semiconductor semantics are typed objects in the data spine — not prompt engineering on a generic schema.
wafer→die→lot→tool genealogy
MSL · hot-lot · qual · ECCN fields
SEMI E10 · JEDEC · SECS/GEM
Your workflows in the training loop
Agents rehearse 8D closure, qual plans and rule-change re-screens on synthetic fabs and OSATs — before they meet your data.
Why now
The people who know how are retiring
Operating knowledge is walking out the door exactly as greenfield plants multiply the engineers who need it on day one.
US semi professionals aged 55+
1 in 3
SIA / Deloitte workforce studies
Knowledge unique to one person
0
Panopto workplace-knowledge survey
Skilled workers needed by 2030
~1M
SEMI industry forecast
Engineer time lost searching
0
McKinsey benchmark
Why a domain-native partner matters now
-
A greenfield plant hires thousandsyoung engineers who have never seen a ramp
-
The answer lives in a retiring headoften at another company entirely
-
The stack is chosen once — at rampwait, and the knowledge has already left the building
Capability, not headcount: playbooks, failure patterns and compliance rules — encoded, cited, queryable from day one.
Simulated console over your own SOPs, work instructions, e-logbooks and NCRs — cited, or an honest abstain.
Enterprise-grade
Credible with your CIO on day one
Designed around the questions security and IT ask before anyone logs in.
Secure by placement
Your walls. Your keys.
Flip for the detailOn-prem or your VPC. Per-plant isolation — yield genealogy never pooled. No training on your data.
Read-first by design
Your ERP stays the record.
Flip for the detailERP, MES, PLM and EDA stay the systems of record. Semi³ reads first — it never writes on its own.
Human-gated writes
Named approver. Exact diff.
Flip for the detailEvery write passes a named approval gate — exact payload, target system, diff. Nothing auto-executes; everything is audit-logged.
Scales with trust
One spine. 37 products.
Flip for the detailOne data spine under 37 products: land with one, attach the next in 60–90 days on the same graph.
Partnership model
Land small. Prove in weeks. Scale on evidence.
no nine-month implementation
no PoC theatre
value first · data last
walk-away clause · in writing
-
Demo
A live synthetic world in your industry: disruption, response, money — in real product UIs.
Your data: none -
Discovery
Your terminology, topology and KPIs shape a synthetic mirror of your operation.
Public + config only -
Workshop
Your team drives. Inject the failures you actually fear; every finding is tracked.
Your data: still none -
Paid pilot
One category or plant, against a signed week-0 baseline and agreed conversion criteria.
Your data enters here -
Production
Certified agents at work; the value ledger keeps score, decision by decision.
Your data — in your twin
Inside the 8-week read-first pilot
Week 0
Connect read-only
ERP extracts plus one MES/test feed, under NDA/DPA with deletion-on-exit. No writes, ever.
Weeks 1–2
Your decision graph resolves
Suppliers, parts, POs and quals become one golden record.
Weeks 3–6
Shadow mode
Agents work your two chosen decisions in parallel — their calls vs your team's.
Week 8
The currency review
Measured evidence on your own data. Continue, expand — or walk away.
Measured value by week 8, or you walk. The pilot is the proof, not the pitch.
The long game
Autonomy is earned, not shipped
Waves overlap — the next starts while the last proves — and autonomy climbs one earned level per wave.
wave 1
wave 2
wave 3
Wave 1 · Months 0–4
The loudest build-phase pains first. Read-first over your ERP — live in 60–90 days, not 9–18 months.
Read-only · recommendations
Wave 2 · Months 3–8
As lots move, the compounding engines attach to the same graph — from yield RCA to die-bank, then Delivery AI.
Write-with-approval · named gates
Wave 3 · Months 6–18
Catalogue products attach in 60–90 days; autonomy graduates per decision-type on eval history — and stays revocable.
Gated autonomy · earned per decision-type
The Owllys story
Shipped software, not slideware
Owllys AI already runs in production across four industries — Semi³ inherits the engines, rebuilt around fab physics.
Paper
Apparel
Logistics
Print
Semiconductors
Purpose-built semiconductor products
0
Seven suites, one platform
Core engines from shipped Owllys software
11 of 12
Only the yield engine was built net-new
DemoForgi code promoted from production
0
13 platform services · ~130 green tests
You're not funding a roadmap. You're inheriting an estate.
proven on a synthetic twin first · zero rows of your data Prove it before you deploy it →
Plain speech
Three things we are not
We could just tell you. We drew it instead.
Not a chatbot wrapper
Deterministic engines compute every number; the model only drafts and cites. Thin evidence — it abstains and routes to a human, never bluffs.
evidence-cited · cite-or-abstainNot a pilot factory
~95% of GenAI pilots show no measurable P&L impact (MIT NANDA). Ours are built to be judged — baseline at week 0, metrics agreed up front.
named metric gates · week-8 walk-awayNot a black box
A named approver sees the exact payload, target and diff before anything executes. Data stays isolated per plant — genealogy never pooled.
named approval gates · per-plant isolation