Skip to main content

We build Swoft with Swoft.
Your enterprise applications follow the same path.

Six of our products already run on this platform: design, business, technical, product, self-service deployment, desktop. The same tech stack delivers your enterprise application in weeks.

  • 2.5 yrs
    Continuous R&D
    A platform built, not a POC
  • 6
    Swoft products
    Run on this same platform
  • 13
    Specialized AI agents
    Constrained by architecture, not by prompt
  • 700 +
    AI tools exposed
    REST + MCP isomorphic by construction
  • 0.2 %
    Technical debt
    SQALE rating A · vs 10-23 % industry avg
  • 0
    Hallucinations in prod
    Non-compliant code rejected at build
01Overview

AI agents generate the code,
the metamodel keeps it in line.

AI code generation is powerful but unpredictable. Our DDD metamodel acts as a structural guardrail: what's not authorized by the domain doesn't pass the build. No drift possible, even under an LLM.

  1. 01AI Designer + business expert

    Domain modeling

    The business is described in the DDD metamodel: domains, bounded contexts, aggregates, events, commands, projections. Each concept becomes typed data the platform understands.

    Artifact
    DDD metamodel · single source of truth
  2. 02AI agents · framed by the metamodel

    Constrained generation

    Agents generate Rust code (domain, handlers, projections, REST, MCP) while honoring the structure defined in the metamodel. Any output that violates a constraint is rejected; the agent can't invent a cross-BC dependency or an out-of-registry endpoint.

    Artifact
    Rust code + isomorphic REST/MCP contracts
  3. 03Quality layer · automatic checks

    Architectural enforcement

    The build fails if a structural rule is violated (Conway, cross-BC dependencies, naming, MongoDB conventions). The quality layer blocks non-compliant code before functional tests even run. No drift possible between design and shipped code.

    Artifact
    Green build or refused build · no in-between
  4. 04AI Runner + platform

    Deployment & operations

    Multi-tenant deployment (DigitalOcean App Platform, on-premise available). Lit UI + REST. AI agents aligned with the domain structure.

    Artifact
    Live application · code transferred
02Architecture

Three layers that reinforce each other.

Each layer has a precise role and depends on the layer below. Together they form a system that understands itself, proves itself, and maintains itself without drifting.

  1. 03 · AGENTIC

    Agentic

    Aligned AI agents · ~700 MCP tools

    Specialized AI agents per business domain, with an architectural scope (not a prompt rule). Conway's Law made executable: each agent knows its domain, its tools, its limits, and can't step outside.

  2. 02 · INTELLIGENCE

    Intelligence

    DDD metamodel · single source of truth

    The DDD metamodel stores APIs, projections, sagas, integration contracts. Single source of truth: the system understands itself and can prove its consistency.

  3. 01 · FOUNDATION

    Foundation

    Event Store · native traceability

    Every action, human or AI, is a stored event, immutable and replayable. Audit proof by construction, non-falsifiable by default.

03How it works

Six technical pillars
that make AI reliable.

Having an AI that generates code isn't a feat. Having an AI whose code respects by design the rules of the business, the law and security — that is. Six structuring choices make it possible.

  • One definition, two surfaces

    The metamodel describes each resource once. The platform exposes it as REST for interfaces and as MCP for AI agents, automatically, without duplication. No drift between the API your users see and the one agents see.

  • Shared identity, distinct roles

    The same person or organization plays multiple roles depending on context: customer in CRM, supplier in procurement, signatory in contracts. Identity stays unique, semantics shift by domain. Native GDPR compliance, no duplicate identifiers.

  • Self-discoverable hypermedia

    Every REST response carries its typed links and actions. AI agents dynamically discover what they can do on each resource — no fragile documentation to maintain, no drifting spec.

  • Decisions tracked as data

    Every decision, human or AI agent, is stored as an immutable event with its author, reasoning, and context. The history is reconstructable end-to-end, without interpretation. Audit by design.

  • Agents reliable by design, not by prompt

    Where most AI agents rely on prompt engineering ("write me an agent that doesn't do X"), ours are constrained by architecture ("structurally, the agent cannot do X"). Reliability isn't a side effect of a good prompt, it's a structural guarantee.

  • Long workflows with automatic compensation

    A business process that crosses 5 steps and 3 services can fail at any moment. Swoft's event-sourced sagas replay inverse actions to bring the system back to a consistent state, with no human intervention. State can't be corrupted by a partial failure.

04Agent representations

Six explicit representations
of your business, shared by all agents.

An LLM alone has an implicit representation of the world, hidden in billions of parameters, non-inspectable. Swoft gives its agents six explicit, structured, verifiable representations. That's what lets multiple agents cooperate without going off the rails.

  • 01Structural

    Business metamodel

    The system knows its own domains, entities, commands, events and their relationships. It's an operational knowledge graph: agents generate code by consulting this representation, they don't guess the structure.

    → The agent knows before acting.

  • 02Temporal

    Event-based memory

    Every action (human or AI) is recorded as a timestamped event with its causal context. Time-travel debugging at any moment T, complete audit trail for GDPR / DORA / NIS2, and replay determinism: AI decisions are stored as data, never recomputed.

    → Nothing is lost, everything is replayable.

  • 03Procedural

    Declarative business logic

    Business rules (conditions, operators, data sources) are declared in configuration and evaluated at runtime. Changing the business logic doesn't require recompilation; a rule change is a data change.

    → The business changes without touching the code.

  • 04Vector

    Knowledge by similarity

    Each agent has a vector embedding knowledge base. Cosine similarity search retrieves similar cases, past decisions, relevant analyses — useful where it's relevant (knowledge retrieval), bounded by agent scopes.

    → The agent recalls what it has already seen.

  • 05Organizational

    Executable Conway's Law

    A routing table maps each business domain to an owner agent and a backup, with confidence thresholds. An agent can't act outside its scope. This is a representation competing multi-agent frameworks (CrewAI, LangGraph, AutoGen) don't have.

    → Each agent knows its limits.

  • 06Visual

    Constrained design system

    Interface components form a bounded visual vocabulary. Static validation rules (ast-grep) reject any UI outside the system. Agents that generate screens pick from this vocabulary; they can't invent free pixels.

    → No visual drift, even from agents.

This combination places Swoft at the frontier of state-of-the-art neurosymbolic AI. In Kautz's taxonomy (AAAI 2022), Swoft is close to type Neuro[Symbolic]: the neural (LLM) is constrained by the symbolic (metamodel), with an additional property Kautz didn't anticipate: the symbolic (Event Store) permanently freezes the neural's decisions, guaranteeing replay determinism. In Sheth, Roy & Gaur's grid (IEEE Intelligent Systems, 2023), Swoft surpasses type 2a federated pipelines (LangChain + solver) while staying below type 2b end-to-end differentiable — a production-ready compromise the literature calls rare. And because the same platform is used to generate our customers' software, delivered applications inherit this same neurosymbolic architecture. Read the full article on the six layers →

05Differentiation

Four properties combined,
no competitor offers them simultaneously.

Where Lovable ships a prototype and LangChain orchestrates agents, Swoft delivers a complete system: domain, events, agents, audit — all by design.

  • Immutable audit trail of AI decisions

    Every agent action is a stored event, replayable 5 years later with its reasoning, model and confidence score. Patented "AI Decision as Data" mechanism.

  • Agents aligned with structure

    An agent's scope is an architectural constraint (Bounded Context), not a prompt rule. The "Compliance" agent structurally cannot modify the "Credit" domain.

  • Guaranteed systemic consistency

    All business domains share a single metamodel. Zero drift, zero inconsistency between design, code, database and API. 0.2 % technical debt (SQALE rating A) vs 10–23 % industry average.

  • Long workflows & compensation

    Cross-domain event-sourced sagas with automatic compensation and approval gates. State can't be corrupted by a partial failure.

06Swoft is built with Swoft

5 internal platforms,
the same tech stack.

All our operations (design, business, tech, product, self-service P2P deployment) run on the same platform. If it stands up to us, it'll stand up to your business.

  • Platform

    SWOFT Core

    Core infrastructure, IAM, Party (shared kernel), AI orchestration, operations, workspace, provisioning. 20 Platform + Party BCs.

  • Platform

    SWOFT Business

    BMC-aligned business operations, analytics, CRM, opportunities, quotes, catalog, resource planning, decisions.

  • Platform

    SWOFT Product

    Design & delivery tooling, GTD, UX design, contracts, product design, code quality, marketplace, scenarios.

  • Platform

    SWOFT Technical

    Technical infrastructure, source control, cloud, mail notification, implementation tracking, packaging, deployment, SSH.

  • Product

    SWOFT Prompt-to-Product

    Self-service SaaS for app generation. The user describes their need in natural language, agents build the full application.

  • Product

    SWOFT Desktop Technical

    Desktop workbench (Tauri + React), domain exploration, screens, SCM, GTD, deployment, agent tooling. Teams' daily driver.

07Compliance

Compliance by design,
not by configuration.

The Event Store structure and metamodel-driven governance natively produce the evidence regulators require.

  • GDPR
    Personal data · multi-tenant via isolated DB
  • NIS2
    Network and information system security
  • DORA
    Digital operational resilience · EU finance
  • Basel IV
    Banking prudential requirements
  • ACPR / ECB
    Financial supervision
  • EASA / FAA
    Aerospace · MRO traceability
  • FDA / EMA
    Clinical trials · pharmacovigilance
  • MiFID II
    Investment services
  • EU AI Act
    In force 02/08/2026 · high-risk obligations

A critical system to deliver?
30 minutes with the CTO.

Derick Schoonbee (30+ years of enterprise architecture) reviews your target architecture and tells you whether Swoft is the right tool.

Request a framing call