The lines to say aloud (tooltips don't reach a projector), the thin ice, and the one-breath answers.
Slide 3 · the thesis — THE line of the meeting
"To be clear: nothing here is built yet. This is the design we want you to greenlight — pressure-tested on paper against 13 products. The ask today is Phase 1 only."
Slide 1 · title
"Two names, one thing: Mimir is the engineering codename for the engine; Modules is the provisional product name. Today I'll say Mimir."
Slide 4 · component map
"The dashed boxes were forced by a design exercise I'll show you on slide 8 — running 13 products through this design on paper."
Slide 5 · capability services
"A signal is a typed fact a stage writes to the case — 'plate matched', 'bumper dented'. Every later stage reads the whole set. This one idea is what keeps the engine generic."
Slide 5 · business model, one breath
"Each of these ships as a sellable API product — priced per session, document, lookup or decision. Clients take one module or the whole stack."
Slide 6 · Priya's case
"Vahan is India's national vehicle registry; RC is the car's registration certificate. The magic beat: the bill lists a new bumper, the photo shows no bumper damage — neither source alone catches it; the fused view does. That fusion is the moat."
Before slides 10–12 · signpost
"Next three slides are for the CTO, the architect, and ML respectively — then we're back together for the plan and the ask."
Slide 8 · back-test
"These numbers are illustrative — a design exercise, not shipped software. The point isn't the 13; it's that the engine-plus-packs split never broke."
"Why build this when vendors exist?" · CTO
We integrate the commodities — OCR, vision/LLM models, workflow engine, model gateway. We build what no vendor sells: the Signal model, the fused image×document checks, and the pack + compliance layer. That composition is the only IP.
"Isn't this a platform before product-market fit?" · CSO
Phase 1 is a product, not a platform — one motor product, end to end. The platform-ness is just disciplined contracts and a CI check. If the product fails, we lost nothing to the platform; if it works, product #2 costs a pack.
"Domain-agnostic — proven?" · Architect
No — design intent. It becomes real the day pack #2 ships with zero engine change (Phase 2). Until then we guard it with the Signal model and a CI check that fails on domain nouns in engine code.
"Is 'agentic' real or fashionable?" · VP Eng
Phase 3, deliberately. The agent's tools are the capability services — they must exist first. Until then, the deterministic DAG does everything we're selling.
"Where do models plug in, and who owns accuracy?" · ML
Everything goes through the Model Gateway (own + third-party, versioned). Review corrections become labels; the eval registry versions models per pack. Designed to publish accuracy — nothing measured in prod today.
"What can I sell first, to whom?" · PM
Single modules self-serve — verification alone, extraction alone, inspection alone — and the composed motor product for insurers. Entitlements make every subset a sellable SKU.
Mimir · speaker crib — pairs with the deck, the agenda and the full Q&A brief.