Run powerful
AI agents at
mass-market
scale.
Bring the use case. We have the full-stack runtime: WhatsApp, voice, persistent memory, EU compliance, and Stripe billing. Built to deliver a genuinely personal experience to every user, at any scale.
Building this yourself means building eight hard things.
Most teams ship two or three correctly and carry the rest as technical debt. Production-grade conversational services are harder than they look.
Every service inherits this from day one.
You define the agent's personality, tools, and prompt templates. The platform owns everything below: wire formats, storage, scheduler, memory, identity, payments, and compliance.
WhatsApp on Meta Cloud API
One dedicated WhatsApp Business Account per service, direct on the Meta Cloud API (not via a BSP). Blue ticks, typing indicators, native push-to-talk voice notes. No shared number, no shared failure mode. Telegram-ready when you need it.
Voice notes, both ways
Inbound speech-to-text via Google Cloud STT (EU endpoint). Outbound text-to-speech via Chirp 3 HD, delivered as native WhatsApp voice notes. Fully async: the platform sends a warm ack immediately, transcribes in the background, and replies when done.
Passwordless phone auth
Six-digit OTP sent via WhatsApp first, SMS fallback on delivery failure. Hashed at rest with a per-row salt. Rate-limited against SMS-pumping attacks. The phone number is the identity. No passwords, no signup form.
Automatic language detection
The platform infers the user's likely language from their phone country code before the first message and greets them accordingly. Mirrors their actual language in every reply. Locale can be locked per user. EN, NO, SV, DA, DE, FR, ES and more supported.
Chat-start attribution
First-touch source captured on every new user: click-to-WhatsApp ad, wa.me link, beta invite, or organic. Stored as a structured event. Admin dashboard shows breakdown by source automatically. Attribution identifiers scrubbed on GDPR erasure.
Third-party account connect
Per-user OAuth callback flow for connecting external accounts: brokers, wearables, any OAuth provider. Tokens stored encrypted. The platform owns the route and CSRF protection; the agent handles the exchange logic.
Mem Palace: durable memory
Everything the user says is stored as verbatim quotes, organized by topic, retrievable as-of any date. Not a summary. Not a context window. A structured, queryable record of what was said and when, across weeks and months.
Semantic search
BGE-M3 embeddings over the full chronicle, stored in pgvector. The agent retrieves contextually relevant memory without scanning the entire history. Session-close summarizer runs as a background timer, keeping recall sharp over long horizons.
Per-product data isolation
Each service runs against its own Postgres database and its own process. A bug, a migration, or a crash in one product cannot leak data into another. Cross-product analytics go through the audit layer only. There is no shared user table.
Conditional SQL triggers
Proactive messages driven by SQL predicates against the live database. When the condition matches, the agent sends. Cron and one-off triggers also supported. No polling service, no external queue. The database is the source of truth.
Approved template management
Meta-approved UTILITY templates for out-of-window delivery. The platform handles the 24h session-window gate, template resolution, and variable filling from the user's own data. One CLI command to register and approve, idempotent, secrets never printed.
Initiative limits
Proactive sends bounded by design. Activity gate, quiet-hours gate, WhatsApp window gate, daily cap per trigger kind. The agent's only autonomous action is a notification. It never acts, decides, or places without the user.
Automated erasure
User requests deletion, confirms with an OTP. The platform erases PII within the hour and removes the account shell within 30 days, retaining only what the law requires. Health and special-category data is deleted outright, not merely anonymised. External OAuth tokens are de-authorised before the state flip.
EU data residency throughout
Secrets, voice transcription, and image analysis all run on EU infrastructure. Personal data and photos do not leave the EU. No user data is used to train any model. GCP europe-west4. Normen-compatible for Norwegian healthcare use cases.
Export + legal pages
User-facing data export (portability) and structured audit log via events. Per-service terms and privacy pages, versioned and admin-releasable. Legal text managed in code with a version gate, so you never accidentally serve a draft or another service's copy.
Stripe built in
Monthly and annual subscriptions, trial state machine, per-turn access gate with brand-voiced block replies. Founder beta invites (operator-issued, 30-day trial). Operator-approved affiliate program with progressive commission. Monthly-equivalent MRR calculated correctly for annual plans.
High-stakes confirmation gate
Any irreversible tool action requires explicit user re-confirmation before execution. The loop intercepts the first call, asks the user, and the tool runs only on the confirmed re-call. Used by Minter for every order. Configurable per service.
Cost-disciplined by design
Per-turn model routing (cheaper model for quick acks, stronger model for real reasoning). Hard cost-brake on metered calls. Per-user daily flood cap, paywall-independent. Per-service economics documented as a CI-gated artifact. Cost is first-class, not an afterthought.
Operator monitoring
Alert engine with WhatsApp delivery to the operator. Debounced so one event alerts once. Per-service checks registered in code. Admin dashboard renders the catalogue automatically. Daily digest available per service. Built on the same channel the users are on.
Image analysis
Inbound images routed through EU Vertex/Gemini, never embedded directly in the main model context. 24h media purge with audit trail. Per-user cap and emergency brake on metered calls. Used by Minter for portfolio screenshot analysis.
Read receipts + typing
Blue ticks on every inbound message, including pending and trial users. Typing indicator on real LLM turns only. URL-bearing replies always delivered as text, never spoken. Single silent path: GDPR-anonymised accounts, which go fully dark.
Three services. One runtime.
Every capability above is live in production, exercised daily. The platform is not a prototype. It is the infrastructure two services run on right now.
AI health coach in WhatsApp. Remembers everything you share (food, sleep, mood, training, energy), connects the patterns, and reaches out first. No new app. The coach that actually knows you.
- Persistent memory across all sessions
- Daily check-ins + proactive follow-ups
- Voice notes in and out
- Food, sleep, mood, training connected
- EU data residency, no training on your data
Calm-trading copilot in WhatsApp. Remembers your investment decisions and why you made them. Never originates a recommendation, never places an order without your confirmation. Flat fee, nothing from your trades.
- Alpaca broker connect via OAuth
- Price watch: notify only, never auto-trade
- Commitment plans + tranche rules
- Portfolio screenshot analysis (EU Vertex)
- Chronicle of every investment decision
AI companion for people with special needs, built for care providers. Norwegian EEA and Normen-compliant by design. The platform's strictest privacy configuration: caregiver visibility, four-scope consent, health-data-grade erasure.
- Caregiver dashboard + shared visibility
- Four-scope consent model
- Normen-compliant (Norwegian care standards)
- Health data deleted outright on request
- B2B to care providers
You bring the soul. The platform does the rest.
SOUL, tools, prompts, schedules. The personality and the use case.
Channel, memory, voice, triggers, GDPR, billing, admin.
WhatsApp. They message a number and start. Nothing to install.
There is no public self-serve API. We are selective about what we build and who we build with.
If you have a use case for a one-to-many conversational agent, write to Daniel. That is where everything starts.
Define your agent
Write the SOUL, the tools, and the prompt templates. Declare which platform capabilities you need in config.py. The platform discovers and wires everything else.
Provision a number
One dedicated WABA per service. The platform wires the webhook, the template index, and the routing. Your number, your billing, your quality rating.
Deploy and observe
Git pull is a deploy. Staged restart per service. Operator monitoring via WhatsApp alerts. Admin console, KPI dashboard, and user management included on day one.
The best way to evaluate a conversational platform is to have a conversation.
We work with founders and operators who want to launch a real service, not a prototype. Write to Daniel. No sales funnel, no deck request. Just a conversation about your use case.
daniel@kvad.studioNo sales funnel. No deck. Just a conversation.