Addi · Architecture Disclosure
Addi is an AI property intelligence platform for coliving and residential investors. We built it as one codebase serving two production lanes, an internal operator dashboard and a public self-serve product, with a first-class AI agent that connects to the platform the same way an engineer's tools do.
The package
This package describes the architecture of Addi at the level where the thinking lives: the boundaries we drew, the failures we designed for, the decisions we made and what they cost us. It stays at that altitude on purpose, because that is where a system's soundness can actually be read.
The boundaries we drew
Where one part of the system ends and the next begins, which side owns what, and how the seams between them are kept honest.
The failures we designed for
The conditions our domain guarantees, wrong records, stale sources, unmapped jurisdictions, jobs that die, agents that drift, and the mechanism that answers each one.
The decisions we made
The forces in tension, the call, and the alternatives we weighed and rejected, including one decision we superseded and kept for the record.
What they cost us
Every mechanism carries a named cost. A quality goal without a mechanism is a slogan, and a mechanism without a cost is marketing.
Where a document stops
Some detail sits below the altitude this package holds. Where a document stops short, it says so in place, so a gap is never something you have to guess at. Notes like this one mark those places throughout the site.Suggested reading order
Traceability closes the loop: check our coverage claims against the package itself. The glossary is there whenever a term needs grounding.
The documents
Architecture
The numbered spine every other artifact cites into: the failure model, the solution strategy, the two lanes, the kernel, the write law, the agent, the compute seam, and how the platform runs.
Decisions
The significant decisions as forces in tension, the call, and its consequences including the negative ones. One superseded decision kept on purpose.
How We Build
The method: outcome to logic to code, the checks between the levels, the two review passes, what done means, and how we build with agents without letting an agent audit itself.
Walkthroughs
Three scenarios with a fictional address: the property's journey, the agent's working morning under the write law, and a bad day where sources fail and records are wrong.
Limitations
The limitations register: what the system does not do yet, which of those are deliberate deferrals, and for each the reasoning and the trigger or plan.
Diagrams
The system views, one abstraction level at a time, each with a declared altitude, a notation key, an element catalog, and what it deliberately excludes.
Glossary
Every domain, platform, doctrine, and agent term the package uses in a specific sense, including the capability-class vocabulary we use in place of data-supplier names.
Traceability
The completeness argument: every question mapped to the sections that answer it, every module traced to the spine and the diagrams, and out-of-scope material as rows, never silence.
The modules
Ten business capabilities and two platform-layer engines. Each unit page carries what it is, its contract shape, and what done means for it.
Intake and Verification
The zero-question front door: an address in, a deal created, a verified identity every downstream module shares.
Buy Box
Explicit acquisition criteria as a living verdict, and the only automatic deal mover in the product.
Pipeline Board
Exactly the published stages and gates, with human approval at every handoff.
Deal Workspace
One deal, composed: artifacts, the findings-to-fields ladder, documents, notes, and tasks.
Reports
A property as a folder of versioned artifacts, every value carrying provenance and confidence.
Underwriting Calculator
One engine serving human and agent identically, every standard assumption labeled and overridable.
Addi Agent Surface
The agent's door: OAuth 2.1 and MCP, the write law, served doctrine, the budget-capped work queue.
Coverage
The honest state of our data assets per geography: covered, partial, or not yet.
Entitlements and Config
The configuration kernel: draft, atomic versioned publish, rollback, fail-closed entitlements.
Signup and Billing
Self-serve signup, idempotent tenant provisioning, and credit-based billing on Stripe.
Analysis Compute Engine
The deterministic engine behind the seam: an address in, four reports and findings out, fail-soft and reproducible.
Zoning Data Factory
The in-house pipeline turning municipal ordinances into canonical, investor-readable zoning records.
The altitude
This package holds to the level of boundaries, guarantees, decisions, and reasoning. Field-level schemas, payloads, formulas, thresholds, prompt and doctrine contents, data source identities, and quantitative measures of our data assets sit below that level and are not part of it. The only outside citations in the package are the public standards we implement, OAuth 2.1 and MCP.