The assistant is the architecture.
Leysco Licia calls the same typed services the screen calls — every action entitlement-checked, correctness-gated, and written to a tamper-evident audit chain. Early access.
repo-governance
repo-governance
repo-governance
AI-native ERP assistant / governed agent platform
SAP Joule · Microsoft Copilot (D365) · Salesforce Agentforce
They bolt an assistant onto a finished suite. Here the assistant and the screen call the same governed services, under one permission model for humans and AI — stated as architecture, chipped Early access. code-inventory
propose_journal_entry → DRAFT
resolve_master_data
ingest_document
… typed tools
Every AI action is a governed tool call.
The gatewayAn MCP tool server exposes the platform as typed, permissioned tools — to Licia and to any external agent. The propose-to-draft pattern: propose_journal_entry returns a DRAFT. The agent proposes; the platform decides. code-inventory
| mandatory fields | checked |
| balance asserted | checked |
| ask-then-act | enforced |
| read-back verify | required |
Zero silent mistakes, by design.
CorrectnessBetween any model and any posting sits DOPS — a deterministic gate: mandatory-field checks, ask-then-act, read-back verification. If the numbers don't verify, nothing posts. code-inventory
Audited like a ledger, because it writes to one.
Trust planePII routing, injection defense, output-integrity checks, cost ceilings — and a hash-chain audit of every AI action. The same discipline the ledger gets, applied to the model. code-inventory
| Provider | Routing | Tier |
|---|---|---|
| Claude | tiered by task | cloud |
| Ollama | self-host default | $0 local |
Your ledger never has to leave the building.
Model freedomClaude tiers routed by task — or local Ollama as a $0 self-host default. Provider-agnostic by architecture, so model choice is a configuration row, not a procurement cycle. code-inventory
The platform under Licia.
In this line| Document ingestion (email/WhatsApp/PDF → governed draft) | Early access | Receipt in, postable draft out. |
| Master Data Resolver | Early access | Money-safe fuzzy entity resolution. |
| Semantic Layer — governed metrics | Early access | One measure, three renderings. |
| MCP Tool Server & agent gateway | Early access | The governed surface itself. |
| Immutable multi-book ledger | Early access | INSERT-only, enforced in persistence. |
| Schema engine (~375 schemas → database) | Early access | Configure-not-code substrate. |
| Authorization — one chain, humans + AI | Early access | RBAC/RLS/FLS/SoD incl. 12 ai.* permissions. |
| 3-tier multi-tenancy | Early access | Tenant-bound isolation and secrets. |
| Licia standalone — bring your own ERP | Early access | Same assistant against SAP B1, ByD, BC, NetSuite. |
| AI eval harness & injection-defense tooling | Designed | Completion-proof for AI features. |
Watch an agent get governed.
Signature · Early access · recorded from development environment · synthetic dataA deterministic replay of a real session shape: the agent lists its tools, asks a governed question, proposes a journal entry — and the platform answers with a gate, not an execution. code-inventory
- agent→ tools/list
- akili← typed tools · scoped by role: semantic.query, propose_journal_entry, …
- agent→ semantic.query { measure: "expenses.total", period: "2026-05", by: "cost_center" }
- akili← 12 rows · governed model · RLS applied for this principal
- agent→ propose_journal_entry { debit: "6200-FUEL", credit: "1000-PETTYCASH", amount: 18500 }
- DOPSGATEmandatory fields OK · balance asserted · read-back required → returned as DRAFT je-2026-000418, not posted
- akili← audit: action hash-chained · cost metered · awaiting human approval
sql-count
The team behind this platform runs 1.66M+ posted invoices in production today — on Leysco100, its sibling house. Licia inherits the team and the domain depth; it claims none of those numbers as its own.
Two houses, independently built. Choose by fit — neither is a stepping stone to the other.
- want to shape an agent-operated ERP from the inside, as a design partner
- need AI that can act — under entitlements, gates and audit — not just chat
- have a data-residency posture that prefers a self-host model tier
- are comfortable with Early access software and honest chips
A chat box beside the ERP, guessing at screens → a governed caller inside it
Part of the house.
RelatedMCP Gateway
The governed tool surface.
Early accessSemantic Layer
Governed metrics under every answer.
Early accessThe Leysco Akili house
The platform Licia is woven into.
Early accessDesigned by the team whose sibling house runs East African distribution floors — informed by that experience, not derived from that code.