Agentix ERP

Problem

Greenfield ERP rebuild with three constraints I gave myself up front:

1. Single shared cluster, no per-tenant database sprawl, full row-level isolation enforced at the data layer rather than only in app code.
2. Custom columns must be queryable with the same indices and predicate pushdown as core columns, not stuck behind a slow EAV table.
3. Tenant onboarding must reach a usable, configured environment from an intake form, without a human integrator on the path.

Approach

Three load-bearing subsystems, designed to compose:

1. Tenant routing. A connection middleware resolves tenant context per request, swaps the active schema, and short-circuits cross-tenant reads at the SQL layer rather than relying only on app-level filters.
2. Zero-migration custom column engine. Tenant-defined columns are stored in a JSON settings registry with ReaderWriterLockSlim concurrency. A typed query builder compiles tenant DSL into parameterized SQL, so a tenant adding a priority_score column gets the same execution plan as a first-class field.
3. 10-step onboarding agent. A four-actor LangGraph-style state machine (Orchestrator, DocumentParser, HelpTextGenerator, Provisioner) reads uploaded Word and PDF intake forms via Amazon Bedrock, with a deterministic rule-based fallback path so the system stays useful when the model is unavailable. Admin credentials are crypto-random and delivered out-of-band via Amazon SES. External approvers receive HMAC-SHA256 signed portal tokens with a 120-hour expiry and a non-reusable nonce.

Stack

• Backend: .NET 10, C#, ASP.NET Core Web API, Entity Framework Core, Hangfire for background jobs.
• Frontend: Angular 17 with signals, TypeScript, no NgRx.
• Data: MySQL 8 with JSON settings tables and generated indexes.
• Cloud: AWS (Bedrock for the document parser, SES for transactional email).
• Containerization: Docker Compose for local and CI parity.
• CI: GitHub Actions pipeline runs xUnit and PowerShell smoke tests.

Outcomes

• 10-step agent ships a configured tenant environment from an intake form, with a deterministic fallback when the model is unavailable.
• Zero-migration custom columns: tenants extend entities at runtime, indexed at the same tier as first-class fields.
• 114 unit tests and 48 end-to-end API tests run on every PR with full tenant isolation, gating merges.
• Portal tokens are signed, time-bound, and single-use, so a leaked approval link cannot be replayed.

Lessons

• A typed query builder over JSON columns is much faster to build than a full ORM extension, and it pays for itself the first time you have to debug a slow tenant query.
• An agentic workflow earns its keep when the agent is constrained to a small registry of vetted operations. The hard work is the registry, not the model.
• Per-tenant feature flags removed entire categories of release risk. A canary tenant pattern caught two regressions before a paying tenant saw them.