NOEVA
IT EN
Log in
Request a demo
Platform evolution

Ten months of development. Eight in production. Zero rewrites.

A platform that landed in production already mature, and has been shipping continuously for eight months without a single architectural rewrite. Inside: the numbers, the principles, the process, the direction.

Months on the platform
18
10 in development + 8 in production, on an architecture never rewritten.
Features shipped
69
Aggregated feat: commits across submodules in the last 8 months live.
Weeks / release
~2
Median release cadence of noeva-shared, the platform's core package.
Mastra workflows
19
Deterministic Zod-typed orchestrations, in production.
Decision principles

What drives every roadmap decision.

Five constraints we apply before writing a single line of code. If a feature doesn't pass these filters, it doesn't enter the plan.

01

Compose, never rewrite

Each release adds a layer on top of the previous one. Day-1 decisions are still valid eight months later: that's the real asset.

02

Per-client data separation from day 1

Organization, workspace, membership and billing live in the data model before any feature. No data separation retrofitted down the road.

03

Open standards only

MCP, OpenAPI, JWT, OAuth, S3-compatible. Multi-provider LLM via AWS Bedrock (Anthropic Claude, OpenAI), data on Supabase Postgres. Customers take their data, integrations and agents elsewhere with zero lock-in.

04

Process safety by default

Timeouts, exponential retries, isolation per worker type, dedicated processes on AWS. Atomic job claim on the Supabase processes table. Long-running tasks never degrade the user's chat experience.

05

Density over surface area

More capability per module, fewer modules. The platform can be understood in half a day, not a quarter of onboarding.

Capability maturity

Six pillars, maturing together.

Each capability grows across the four phases, from seed to composable. Shows the platform is engineered by pillars, not opportunistically per feature.

Knowledge base management
Sep
Jan
Mar
May
Agents and Super Agents NOEVA
Sep
Jan
Mar
May
Deterministic workflows
Sep
Jan
Mar
May
External integrations
Sep
Jan
Mar
May
Billing and tenancy
Sep
Jan
Mar
May
Observability and telemetry
Sep
Jan
Mar
May
Evolution timeline

Four phases, one direction, no rewrites.

Each phase built on top of the previous, with a clear context, a key architectural decision, and a concrete metric anchoring it to reality.

  1. 01
    Sep 2025
    Proof point
    32 tables, 8 worker functions, first workspace in production.
    The fundamentals. Identity, data, first tools.

    Foundation

    Before any AI feature, we needed a solid platform with data separated per client: identity, workspace, file primitives and an initial RAG pipeline to anchor everything else.

    Shipped
    • Core API with per-client data separation: organization, workspace, membership, roles and invites.
    • Supabase data schema with 32 relational tables and RLS by default.
    • Initial RAG pipeline: ingestion, AWS Bedrock embeddings and semantic search on Supabase pgvector.
    • Supabase Auth + cross-domain JWT for the app, client and future MCP servers.
    • First 8 worker functions: classification, summary, embeddings, Firecrawl-based web crawling.
    Key decision

    Multi-tenancy in the data model before any feature: no retrofitting later, no painful migrations.

  2. 02
    Jan 2026
    Proof point
    18 worker functions (api), 1 heavy dedicated to EXTRACT_TEXT.
    Documents become queryable knowledge.

    Workspace & AI Ready

    Once workspaces worked, the real blocker was turning uploaded documents into AI-ready knowledge without manual intervention. The AI Ready concept was born here.

    Shipped
    • Automatic AI Ready pipeline on upload: extraction, classification, embeddings.
    • Hybrid semantic + keyword search with workspace-aware re-ranking.
    • Multi-worker orchestration on AWS: 5 api + 3 heavy instances with isolation per workload type.
    • First specialized Mastra agents: classification, extraction, search.
    • Workspace billing model + token management per organization.
    Key decision

    Per-type worker whitelist with runtime isolation: heavy tasks never degrade the synchronous chat experience.

  3. 03
    Mar 2026
    Proof point
    19 workflows in production, 8 external providers integrated.
    AI enters daily processes.

    Operational AI

    With workspaces stable and atomic agents in production, the next step was wiring AI into customers' real processes: CRM, HR, communications, scheduled jobs.

    Shipped
    • Deterministic Zod-typed Mastra workflows for CRM, time tracking and HR.
    • WhatsApp Super Agent + customer entry routing over WhatsApp Business API.
    • Scheduled jobs runtime with dispatch to worker or workflow.
    • Bidirectional Twilio voice routes for voice agents.
    • Embeddable Agents UI v1 via iframe + postMessage protocol.
    Key decision

    Deterministic workflows alongside agentic agents: every critical process has a DAG-typed variant without LLM creativity.

  4. 04
    May 2026
    Proof point
    2 MCP servers, 37 agents, 109 tools, 19 workflows.
    A platform that composes, measures and bills itself.

    Marketplace & Platform

    The last macro layer turns NOEVA from a vertical product into a composable platform: catalog of agents and workflows, per-workspace billing, distributed MCP servers and enterprise governance.

    Shipped
    • Marketplace catalog of agents and tools, browsable by domain.
    • 37 production agents registered in the Mastra runtime.
    • 109 unique atomic tools, organized by domain (CRM, RAG, SEO, scheduled).
    • Distributed MCP servers: noeva-mcp-core + noeva-mcp-crm over public npm (@noeva).
    • Personal API keys (noeva_sk_) for user-scoped external integrations.
    Key decision

    MCP as the official interop protocol: the platform exposes itself to external clients without inventing proprietary contracts.

In flight

What we're shipping this quarter.

No exact dates, but six concrete worksites open right now, each with a clear reason to exist. The sign that the platform is alive, every week.

  1. Now · in flight

    noeva-mcp-agents (third MCP server)

    Exposes the 37 Mastra agents as standard MCP tools, authenticated via Personal API key: enables external orchestration from Claude, Cursor and any MCP-compatible client.

  2. Now · in flight

    Personal API keys (noeva_sk_)

    User-scoped tokens with extensible scopes: a prerequisite for external MCP, custom integrations and on-behalf-of automation without exposing workspace credentials.

  3. Now · in flight

    Embeddable Agents UI v2

    Iframe + postMessage protocol matured: theme sync, navigation state, auth refresh, to integrate the agents chat into any white-label portal.

  4. Now · in flight

    Workspace-builder agent

    An onboarding agent that produces the initial workspace knowledge base from a structured conversation: shortens new-customer time-to-value.

  5. Next · planned

    Agent observability + evals

    Per-run agent tracing, per-model cost attribution, continuous eval sets: enterprise prerequisite to put LLMs in critical flows with accountability.

  6. Next · planned

    Marketplace billing & revenue share

    Catalog + per-workspace billing + revenue share: turns the marketplace from a showcase into a composable economy for partners and agent creators.

Strategic themes

Three directions we're investing in.

No dates, no promises: the direction that drives our architecture and investment decisions. What you'll see in six or twelve months starts here.

01

AI-native compliance

Granular audit trail, RBAC on every resource, EU data residency, ISO 27001/27017 alignment. Compliance isn't bolted on later: it's engineered from the start.

02

Marketplace economics

Catalog, per-workspace billing, revenue share, cost attribution per token and per model. A platform that composes, measures and bills itself.

03

Agent observability

Tracing for every agent run, continuous evaluations, per-workspace cost attribution. LLMs in production aren't a black box: they're a measurable system.

no dates · no promises · just direction
How we ship

How we deliver all of this, every week.

A roadmap without a process is a list of promises. Four practices we apply to every release, from code review to runtime isolation.

Continuous cadence

Small, frequent releases, continuous integration on main, fast rollback. Moving quickly is a property of the process, not an aspiration.

Review before merge

Automatic code review and security review on every PR. No code in main without multiple eyes, no shortcuts on critical features.

Isolation by default

Worker types dedicated by purpose (api / heavy) on AWS, scheduled jobs separate from the synchronous pipeline. The Supabase processes table acts as the atomic broker for job claiming. A slow pipeline never slows down the others.

Live telemetry

Sentry + Supabase metrics + structured logging on AWS CloudWatch; Cloudflare Analytics at the public edge. We know what the platform is doing in real time, not just when something breaks.

Release notes

Want to see every release in detail?

The public changelog tells you what shipped to production, version by version.

Open the Changelog →