Ghasi AI Intake OS
In development · pre-launch
- · AI-first by design
- · Offline-first by design
Domain-agnostic, multi-tenant, AI-first intake operating system for government and regulated industries. One engine, two surfaces, many domain packs.
- Period
- 2025 — present
- Domain
- GovTech / Regulated B2B + B2C
- Stack
- API Gateway · IAM
- Per-surface BFFs
- Relational + Vector
- Event Bus
- AI Gateway · Multi-LLM
- DDD · Outbox
Public-sector intake — visas, permits, claims, registrations — runs on spreadsheets, paper, and bespoke portals. Each agency rebuilds the same plumbing: forms, eligibility rules, document handling, identity, audit.
Intake OS replaces that plumbing with a single multi-tenant engine and domain packs that can be swapped per agency.
The platform is AI-first and offline-first by design, not by integration. AI-assisted extraction, eligibility assessment, and document analysis run through a single AI Gateway with provenance metadata attached to every output. Field and provincial offices work offline with a local outbox that reconciles when connectivity is restored — a hard requirement for government programs operating in regions with unreliable infrastructure.
The layers, ports, and integration patterns are real — drawn from the live architecture baseline. Service names, identity providers, and external vendors are anonymized or labelled by role so the diagram reads as the posture, not as a deployment manifest. Hover any layer label on the left for context; right-edge annotations call out the load-bearing decisions.
— 01 —
TypeScript-first stack across services and clients.
CostCeiling on CPU-bound workloads.
WinOne language, one type system, one toolchain.
— 02 —
Clean Architecture / DDD with the domain layer in pure TS.
CostMore files, more interfaces.
WinEvery adapter (DB, queue, LLM) is replaceable without touching domain.
— 03 —
Multi-tenancy enforced at three layers — domain, row-level security at the relational layer, tenant header on every event envelope.
CostTenant context plumbing in every request.
WinNo cross-tenant data leak is structurally possible.
— 04 —
LLM provider abstracted behind a single AI Gateway port.
CostA small adapter to maintain per provider.
WinThe business decision of "which model" doesn't ripple into application code.
— 05 —
Outbox pattern for every state-changing event.
CostExtra table per BC, relay process.
WinExactly-once-effective delivery without distributed transactions.
- First bounded context (IAM) wired with full DDD layout, OTel spans, OpenAPI, signed container, CI gates green.
- Spec-first: every BC has an implementation blueprint that CI verifies against the live OpenAPI (drift detection on every PR).
- Sprint-1 wave: full IAM aggregates, ethical-wall checker, outbox relay.
- Approaching pre-launch deployment for first pilot agency.
Founder, lead architect, and primary contributor. I authored the spec, the architecture baseline, the BC map, the implementation blueprint template, and the IAM service end-to-end.
The provider port abstractions in V1 are slightly leaky — the LLM port exposes streaming primitives that should live a layer up in the application service. Refactor planned for Sprint 2.