Shopping cart
Your cart empty!
// Capabilities
The shared engine. What we ship in every engagement.
Applied AI on a mature stack: the same engine that powers our flagship WhatsApp product runs every bespoke solution. Production-grade defaults, audit-friendly choices, no homegrown frameworks.
What we ship
Four capabilities, production-tested.
Each capability is what the engine actually does today — built once, applied per engagement.
Conversational agents
LLM + tool use + memory
Free-LLM-first runtime with deterministic rails on critical-action intents. The agent decides when to talk and when to call a tool; memory carries the conversation across turns without prompt-stuffing.
Example — a clinic bot that books an appointment on the practice's real Google Calendar, then sends the confirmation back over WhatsApp.
- Claude
- GPT-4o
- Tool use
- Context cache
RAG over custom knowledge bases
PDFs, sheets, URLs, Q&A
Documents and structured content are chunked, embedded, and stored in a vector store. Retrieval uses a cross-encoder reranker; citations are operator-visible. Knowledge bases are scoped per workspace — no cross-tenant leakage.
Example — uploading a 40-page service catalog and a Q&A sheet gives the agent grounded answers from day one, no model fine-tuning needed.
- text-embedding-3-large
- Qdrant
- Reranker
- Per-workspace scoping
Integrations with external systems
Calendars, ERPs, CRMs, catalogs
Pluggable tool adapters with explicit auth patterns. Each adapter ships with retry policies, idempotency keys where the upstream supports them, and structured error surfacing back to the agent.
Example — Google Calendar read/write (OAuth), HubSpot contact upsert (Bearer), inventory lookup against a SQL replica (API key).
- OAuth 2.0
- Bearer
- API key
- Basic auth
Internal-process automation
Team assistants, document processing, AI search
The same engine powers internal-facing surfaces: assistants that read a team's private corpus, document classifiers that route incoming files, AI-powered search across years of operational history.
Example — a back-office assistant that triages incoming support tickets against the policy KB and routes them to the right queue with a draft response.
- Slack adapter
- Document classifier
- Semantic search
- Operator-only citations
Stack
Architecture
The engine, in 30 seconds.
Same pipeline runs WhatsApp PyME bots and bespoke mid-market engagements. The block on the left changes per channel; the rest is shared.
Channel
WhatsApp Cloud API today. Web chat, voice, email pluggable as adapters.
Intent classifier
Routes critical-action intents into deterministic rails; everything else flows into the free-LLM lane.
Flow engine + Free-LLM + RAG
Rail flows for book / order / pay / escalate. Free-LLM with retrieval for everything else.
Tools
Calendar, ERP, CRM, catalog, email — explicit adapters with typed schemas.
Action
The action that defines success — booked, captured, routed, paid.
Each node is observable: every turn carries the intent, the retrieved chunks, the tool calls, and the final action — visible to operators today and to clients in v1.1.
Coming in v1.1
What we are not promising yet.
We name what is still in the lab so the contract stays honest. Both items below are in active build but not exposed to clients yet.
v1.1
Evals exposed to the client
Today we run evals internally per engagement: regression suites against the real-action metric, holdouts per release. The client-facing eval dashboard (read-only access to suites, history, regressions) ships in v1.1.
v1.1
Client-visible observability
Operator-side observability (every turn, every tool call, every retrieval) is live today. The client-side cut — same data, scoped to their workspace, with audit-grade export — ships in v1.1.
Want to see how we would apply this to your case?
Thirty minutes on a call, no slides. Bring the action that needs to fire, the system it lives in, and the volume — we will sketch the engagement on the call.