We build an AI framework in which an IT team delivers changes faster than the business can think them up: a business developer states the change, agents write code against contracts, QA verifies before deploy, DevOps ships it as Temporal workflows. FinTech today, ERP and e-commerce next.
Threads of the registry woven into an executable process
AxonWeaver does not ship as an empty framework. The process templates of a lending business are already built in: designed by a team that has delivered end-to-end fintech products on many markets, from consumer cards in Mexico to point-of-sale financing in Uzbekistan. You customise verified processes instead of inventing them from scratch.
Payday loans, BNPL, installment, collateralised lending, credit cards, credit lines, earned-wage access, E-POS financing. Product parameters, quote logic and billing rules live in the registry, not in a consultant's head.
Onboarding and OTP, periodic and triggered KYC, client block and unblock, loan application sagas, contract → activation → disbursement, servicing and billing, support and restructuring, marketing and compliance. Each is an executable orchestration, not a diagram.
Customer journey, KYC and compliance, credit underwriting, settlements, payments, disbursement, collections, notifications, documents, product configuration, finance reporting, support operations. Every node has validated input and output contracts.
Categories of external APIs the domain needs: payments, identity and KYC, banking, communications, data and scoring, partners and leads. Credit decisions run in Axon Risk: strategy graphs, A/B and shadow versions, bureau and ML sources.
The result: an IT team starts from a populated registry. Changes are customisation of verified templates, and the generated code inherits the domain logic rather than reinventing it.
The FinTech domain in AxonWeaver covers a lending product as a whole, from the first contact with a customer to contract closure and reporting. Templates ship for payday loans, BNPL, installment, collateralised lending, credit cards and credit lines. The customer journey opens with onboarding and verification (OTP, KYC, compliance), passes through underwriting with a credit decision in Axon Risk and a manual review wherever a person is required, then contract, mandate and disbursement through bank and wallet rails. After disbursement the platform runs servicing: billing and schedules, payments and their allocation, restructuring, notifications, collections in soft, hard and legal stages, financial reporting and postings export. All of these processes share one registry model, execute as Temporal workflows and change along one and the same path: the four roles described below.
One change, four hands, one deployable package
States what has to change. Fills in structured questionnaires across six areas: screens, integrations, documents, workstations, product catalogue, communications. Fixes the Mission Statement and the mandatory artefacts. The output is a deterministic change specification, not free text.
Turns the specification into a registry model. Entities with typed attributes, nodes with input and output contracts, process graphs, BPMN of business processes, coordinator rules and external APIs. An AI agent generates code from this model through MCP; validation errors go back to the agent, not to a person.
Verifies the change before deploy. Process dry-runs with automatically generated negative scenarios, runs in mock, live and mixed modes, an external API sandbox, reference product and billing cases, code checked against the source model. A package does not move on without a green report.
Ships and observes. A process package is deployed into the project's Temporal namespace with its own data schema, schedules, webhooks and human tasks. Project documentation is built automatically, rollback is the same package, execution is visible in Temporal UI and the audit log.
The agent does not write whatever it likes: it calls registry tools and the platform validates every artefact against its contract. Code is stored with a sha256, the process manifest describes starts, data and tasks. Everything the agent does is journalled per operator session.
A Temporal worker and namespace per project, Schedules from the manifest, waiting for a human decision through signals with SLA escalation, a separate data schema, code export and project documentation in mkdocs. Nothing needs to be “finished for production”.
An open stack; all generated code and infrastructure are handed over to the customer. The primary agent is Claude Code, yet the platform works with any MCP client, and models for a closed perimeter can run locally. Domain knowledge is formalised in the registry, not in a contractor's head.
The platform consists of independent services with a shared data model. Each is available to AI agents through its own MCP server.
One registry in the middle, four products around it
Registry of entities, nodes, processes and business processes; BP Designer, signal coordinator, packages, deploy, project documentation, role workstations.
Decision engine: strategy graphs, draft / A/B / shadow versions, registry of Python procedures and external sources, MLflow, execution journal, replay and regression.
Knowledge base over the registry, strategies and documentation: hybrid search, reranker, chat with sources. Next to it an ETL and BI loop on Dagster, DuckDB and Superset with single sign-on through Keycloak.
Customer workstations are assembled by the agent in FlutterFlow: registry methods are registered as APIs and bound to screens; screens are checked against the Figma designs; checks in mock and live modes.
A payday loan product for South Africa with local rules: VAT, DebiCheck, s.126 payment allocation, the in duplum rule. Eight business processes, from product quote to reporting, went requirements → registry → agent → package → Temporal. Thirty days passed from the brief to the launch into pilot operation.
Three rings of protection and an eye that looks from outside
Every public entry point passes TLS at Caddy and an OIDC login in Keycloak with the roles admin, analyst and viewer. Calls from AI agents carry a Keycloak token or a service bearer and are bound to an operator session; every tool call is journalled and rate-limited. Service ports listen on loopback only.
A single login for the knowledge base, BI, orchestration, the workbench and the admin: OIDC through oauth2-proxy, roles read from the access token, service accounts on client credentials, JWT audience checks on every MCP server.
Caddy with Let's Encrypt and HSTS on every subdomain, secure cookies, direct service ports on loopback or behind a proxy.
A bearer at the transport and an operator session inside every tool, an idle timeout, rate limits per session and per bearer, an SSRF guard and import deny-lists in the code sandbox.
Admin history, LLM audit events, the execution journal of Axon Risk, generated code stored with a checksum.
Heavy components live on a separate monitoring machine; the other hosts carry only light agents. Part of the checks run from outside, over the network, the way an attacker would look at us. Four tasks, all open source:
Caddy, oauth2-proxy and Keycloak at the door; CrowdSec reads the logs and bans an address on behaviour: password spraying, path scanning. Automatic, a shared community block-list on top.
Wazuh agents on every host track file integrity, logs and anomalies and send events to the security center. Alert, reviewed the same day.
Wazuh compares installed packages with CVE feeds; Greenbone scans every host from outside, the way an attacker sees it.
A single event journal from all hosts in the Wazuh indexer, entry through a WireGuard bastion, dashboards reachable only through an ssh tunnel.
Automatic actions are deliberately few: a banned address costs nothing, a killed process on production costs downtime. Everything else goes to a person, through one channel for all findings.
What it does not cover: a correct login with a stolen password (that is the job of a second factor and short token lifetimes), the content of TLS traffic, and application logic, which stays with the role model and review.
No proprietary engines: what you can hire for on the market and maintain yourselves.
Loan management, risk, CRM and collection, data warehouses: from concept to production in high-load brokerage and lending platforms.
The platform is deployed in your perimeter, the team goes through knowledge transfer and then designs and generates on its own. A fixed monthly fee.
The AxonWeaver team takes systems analysis, architecture supervision, QA and DevOps; you focus on the product.
Early access to the new domains in exchange for real processes and feedback.
Demo access, an architecture walkthrough and a timeline estimate for your domain.