← Mimir docs

Mimir — one case, end to end

Read each row left→right: what happens · the component that does it (built once) · what the pack supplies (per domain).

Platform — built once, domain-agnostic Pack — per-domain configuration Cross-cutting — always on
The example: Priya dents her car and files a motor insurance claim. The insurer runs it through a product built on Mimir (pack = Motor). Watch one case flow through the engine — the middle column never changes between products; only the right column does.
What happens (the story)
Mimir component
Pack supplies
1
Collect the evidenceInsurer texts Priya a link. From her phone she shoots the car and uploads her RC + repair estimate. Blurry shots are rejected on the spot.quality check: sync

Does it

Capture-Link + Capture-Quality
No-install link (SMS/WhatsApp/email), guided shots, re-shoot prompts, consent recorded.

Pack

Which angles & documents to ask for · consent copy · language (Hindi/English).
2
Open the caseA case is created: consent stored, data pinned to India, everything from here is logged.async from here

Does it

Case Service
Event-sourced unit of work — the log is the audit trail.

Pack

Residency = India · retention window.
3
Understand itThe documents become fields (policy no., dates, amounts). The photos become findings: front bumper — dented — moderate.

Does it

Extract + Vision
Raw → typed signals on the case. Both run via the Model Gateway.

Pack

Document schemas · the motor feature set (parts, damage, severity) · model choice.
4
Check it's realPlate & chassis in the photo are matched against Vahan; owner against the RC/PAN.

Does it

Verify
Cross-checks against pluggable sources of truth; async, cached, fails soft.

Pack

Which registries count: RC/Vahan, DL, PAN, GST.
5
Cost itThe dented bumper maps to parts + labour → an itemised repair estimate.

Does it

Estimate + Reference Data
Feature signals × priced catalogs → costed line items.

Pack

Parts catalog · labour rates (licensed, versioned).
6
Score the riskFused view: the estimate claims parts with no visible damage, and the policy lapsed before the loss. Risk = 78, high.fused check = the moat

Does it

Detect
Reads the whole signal set — images, documents, records — and flags with the exact figures behind each.

Pack

Risk rules & weights (IRDAI-mapped).
7
DecideClean claims auto-clear. Priya's scores high, so it routes to an investigator — with the reasons attached. Never auto-cleared above the legal cap.human owns the call

Does it

Decide (Flow)
Thresholds + hard limits. Pauses durably for human review.

Pack

Thresholds · routing · legal auto-approve cap.
8
Deliver & learnInsurer gets a JSON result, a PDF report and a webhook. The investigator's verdict is recorded — and trains the engine.

Does it

Webhook/Reports + Review & Feedback
JSON + branded PDF + audit entry; corrections become labels.

Pack

Report template · branding.

Always on, underneath every step

Orchestration — runs the steps (fixed DAG or an AI agent) Model Gateway — all models, compliance-aware routing Compliance — residency · consent · retention · audit (a USP) Entitlements & Metering — who can use what · usage billing

Same engine — swap the pack, get a different product

Steps 1–8 and the whole middle column stay identical. Point it at a new domain by changing only the right column.

Lending — collateral checkVision grades the vehicle · Verify checks loan-on-RC · Estimate values it. Pack: bureau sources, valuation tables.
KYC / onboardingExtract reads IDs · Verify matches registries · Detect spots forgery. Pack: ID schemas, forgery rules. No Vision, no Estimate.
Health claim checksExtract reads bills · Detect spots up-coding. Pack: procedure price lists, health-data compliance regime.
Fleet — before/afterSame capture, new question: "what changed on return?" Pack: condition schema, compare mode on.

Mimir · the Modules intelligence engine — platform architecture at a glance. Full detail in the docs set (README → 01–06).