Pillar 02 · Built by Orokii

A conversational AI that ships in production, not in a demo.

The Kora assistant — Claude-powered, multilingual, slot-filling, compliance-aware — already runs in Stratum Remit, Sikama, and LuckyCat. We license the runtime and the prompt pack so your bank, wallet, or lending app can ship the same shape of experience.

Where the market is

Every fintech wants conversational AI. Most are still on rules-based bots.

Established

Bretton AI, Sardine, Coris, the YC compliance cohort. Strong on back-office analyst augmentation, sanctions triage, KYB. Less on consumer-facing turn handling.

Underserved

Production-grade customer-facing conversational AI for regulated fintech. Multilingual support for African and Latin American languages. Compliance-slot patterns out of the box.

Where we sit

Customer-facing, multilingual, compliance-aware, and already in production across three tenants. Sell the runtime, not the demo.

What's included

Runtime, prompts, workers, and integration support.

Hybrid intent parser

Regex + keyword fast-path resolves ~95% of phrasings without an LLM call. Claude tool_use fallback handles the rest. Token spend stays bounded; latency stays under a second on the common path.

Production · Stratum, Sikama, LuckyCat

Multilingual

English, Spanish, French, Yoruba, Swahili out of the box. Language detection per turn. New languages added by extending the prompt pack — not a code release.

Production · 5 languages live

Slot-filling state machine

Fourteen intent types across seven conversational states (amount, recipient, wallet, reason, source of funds, relationship, completed). Domain-pluggable so banking, lending, insurance, and wallet domains each define their own intent set without forking the runtime.

Production (remittance) · pluggable framework needs decoupling

Smart reminders and proactive surfaces

FX rate alerts, inactivity reminders, birthday and special-date detection, recurring transfer scheduling. Each one a discrete background worker; turn on the ones you want.

Production

Compliance-aware turn handling

Built-in slots for source-of-funds, transfer reason, beneficiary relationship — the regulator-facing checks that consumer fintech apps skip and that breaks them at scale.

Production

Hosted control plane (optional)

Prompt management, conversation analytics, A/B testing, per-tenant prompt isolation. Skip if you want to self-host the prompts; turn on if you want the operational tooling.

New build · ~4-6 weeks

Who we work with

Three places the Kora pattern fits cleanly today.

Neobanks adding conversational UX

Plug Kora into the existing app. Replace 4–6 tap-to-navigate flows with a single chat surface. Compliance slot handling already wired in.

Lending and insurance apps

Loan status, repayment schedules, claim updates, document collection. Conversational replaces the typical FAQ + IVR + chatbot trio.

Wallets and CBA installations

Balance, transfer, statement queries — without writing your own retrieval-augmented generation. Cross-pillar pull-through from Kora CBA deployments.

How we work together

You own the customer. We license the runtime.

You bring

  • Your customer app and identity stack
  • Your domain — the intents, slot definitions, and tool calls specific to your product
  • Your Anthropic API key (or your hosted Claude integration)
  • Your data — conversation logs stay on your infrastructure

We bring

  • The runtime — intent parser, state machine, slot fillers, background workers
  • Multilingual prompt pack (en/es/fr/yo/sw, extensible)
  • Reference integration with your existing app
  • Optional hosted control plane for prompts and analytics
  • Production proof — the same code runs in Stratum, Sikama, and LuckyCat today
  • Six years of compliance-tuning experience for regulated conversational flows

Commercial model

Delivery engagement first. Optional SaaS layer after.

Delivery engagement

$75k – $150k

one-time integration

Wire the runtime into your existing app. Configure intents and slots for your domain. Tune the multilingual prompt pack. Hand off the runbook.

Optional hosted control plane

$3k – $10k

per month · scales with conversation volume

Prompt management UI, conversation analytics, A/B testing, per-tenant prompt isolation. Skip if you self-host the prompts.

LLM token cost (Anthropic API) is separate and billed to your Anthropic account. Numbers above are guidance; final scope and price are set in the engagement letter.

Frequently asked

What partners ask before signing.

Why Claude specifically?

Stratum and the whitelabel tenants run on Claude Sonnet 4 today. The intent-parsing pattern is model-agnostic, but every tuned prompt and every production conversation log is against Claude. Migrating to a different provider is achievable but is engineering work that needs scoping. We can talk through it during the call.

Will this work for our non-remittance domain?

The Kora runtime as it ships today is remittance-specific in its intent enum and slot vocabulary. The architecture supports a domain-pluggable refactor (estimated 3 months, AI-augmented) so banking, wallet, lending, and insurance domains can each define their own intents. Whether you wait for that or pay for the decoupling as part of a delivery engagement is a scoping conversation.

How does this play with our existing chatbot or LLM stack?

Kora is most useful where the existing surface is a rules-based bot, an FAQ widget, or nothing. If you have a serious LLM stack already, the win is the slot-filling state machine and the multilingual prompt pack — those plug in as a layer rather than a replacement.

What about LLM cost?

The hybrid parser is the answer to this. Regex fast-path resolves the majority of common phrasings with no LLM call. The state machine reuses cached context across turns. Reference deployments at Stratum, Sikama, LuckyCat run inside reasonable per-conversation cost envelopes. Final numbers depend on your traffic mix; we benchmark during scoping.

Where does the conversation data live?

On your infrastructure by default. Self-host is the standard delivery shape. The hosted control plane (optional) is for prompt + analytics tooling — your conversation logs stay yours.

How long until first deployment?

For a remittance or remittance-shaped product: 6–10 weeks from contract to live. For a new domain that needs the runtime decoupled: add the 3-month framework build. The decoupling work is reusable across future engagements so the second customer in a new domain is on the shorter timeline.

Ready to evaluate?

A 30-minute call covers your domain, your existing AI surface (if any), your LLM provider preference, and rough deployment timeline. We tell you whether the fit is real.