LeaseFlow: Codex for Everyone

Turn messy landlord documents and real user feedback into governed leasing workflows that always use the latest approved information.

Comment

Inspiration

This project began as a conversation, not a codebase.

Commercial leasing teams receive updated flyers, floor plans, area workbooks, contracts, calls, and email requests from many owners and buildings. The documents look familiar even when a number, available floor, incentive, or plan has changed. A team member can spend more time reconstructing what is current than responding to the broker or preparing the weekly owner meeting.

I started with the work itself: the documents, the handoffs, and the corrections from the person who actually does the job. I used Codex to turn that lived workflow into an information architecture, a data authority model, a working Admin Web and Mobile experience, and a repeatable test path.

The result is LeaseFlow. The broader idea is Codex for Everyone: a domain expert should be able to describe real work in ordinary language, show the artifacts, challenge the design, and collaborate with Codex until the bottleneck becomes a usable, verifiable workflow.

What it does

LeaseFlow connects five parts of leasing operations.

  1. Register and understand landlord source material. A designated user chooses the building and source type, uploads a document, and reviews the extracted content. Text-based PDF, DOCX, and XLSX files can become structured review material. Image-only or unsupported drawings stop for manual review instead of silently becoming official data. Large landlord documents are analyzed rather than rejected only because of file size.
  2. Confirm and publish current building information. Extracted changes remain proposals. A Data Steward confirms them and a Senior Reviewer publishes them. Versioned fields such as available area, floors, incentives, parking, availability dates, and floor plans retain their source and history. A superseded floor plan cannot re-enter a current external package.
  3. Use the same current information across Web and Mobile. Team members can find a building, inspect its latest operational facts, prepare a customer package, and see the original source link without learning internal state names or program concepts.
  4. Automate weekly work by landlord. A manager groups assigned buildings under the correct landlord, configures the weekly meeting schedule, required sections, recipients, and approver, then prepares separate building-specific reports from the week's approved information and activity.
  5. Keep consequential actions human-owned. AI output is a candidate, draft, or patch. Deterministic code controls publication, authorization, current-version selection, recipient groups, and delivery gates. External packages and reports require a named human approval. The demo records sandbox deliveries only.

The verified path shows a 5F record changing from 300 py to 200 py, rent-free from three months to two, supported parking from three spaces to two, and floor plan v1 to v2. After publication, both Web and Mobile use v2 and exclude v1.

How we built it

LeaseFlow is an npm-workspace TypeScript monorepo.

  • Next.js provides the Admin Web, shared Web workflows, and demo API.
  • Expo and React Native provide the mobile operations experience and Mobile Web test surface.
  • A shared domain package owns roles, publication rules, authorization, current-version selection, recipient calculation, and send gates.
  • Zod validates source-extraction, request-extraction, and weekly-report patch candidates.
  • A resettable single-process JSON store makes the synthetic judge path reproducible.
  • Vitest, Node tests, TypeScript strict checks, build checks, and browser QA protect the end-to-end workflow.

The credential-free demo uses deterministic synthetic candidates under the same validation contracts as the optional server-side OpenAI adapter. The adapter is designed for bounded GPT-5.6 proposals and sends store: false, but the final verified product path does not claim a live external model call, Outlook delivery, SSO, carrier integration, or production datastore.

How we used Codex and GPT-5.6

GPT-5.6 powered the Codex collaboration used to build and refine the project. Codex did more than generate a first implementation.

  • It read the product specification, workflow notes, and source-document patterns.
  • It translated natural-language corrections into a shared data model, state transitions, navigation, labels, and acceptance tests.
  • It responded to annotated browser feedback such as removing redundant explanations, showing current work instead of feature tiles, listing all assigned buildings, adding return navigation, and aligning status indicators consistently.
  • It helped test document categories and convert failures into product rules. For example, a 20MB upload cap was removed after the user explained that landlord source material must be analyzed regardless of size; the revised path successfully processed a 62MB, 358-page synthetic test document into human review.
  • It created evidence for the complete upload → review → publish → Web/Mobile synchronization path and prepared a remote evaluation route for the real user.

The human remained the source of domain truth. Codex converted that knowledge into inspectable product behavior and tests. The repository includes a thin codex-for-everyone skill that distills this collaboration loop so another person can begin from their own work instead of from software vocabulary.

Challenges we ran into

  • The first designs looked like feature dashboards rather than the user's actual day. Repeated browser feedback forced the Home experience to center current work, latest building changes, and weekly-report progress.
  • A fluent extraction is not the same as an official fact. We had to separate source registration, AI proposals, Data Steward confirmation, Senior publication, and external use.
  • Commercial documents vary widely. A text-heavy flyer, area workbook, legal document, image-based perspective, drawing file, and very large portfolio PDF need different review boundaries.
  • A weekly owner meeting may cover several buildings, but each external report must remain building-specific and use the configured recipient and approval authority.
  • A remote real-user test exposed deployment truth: a single-process JSON demo can be shared through a temporary tunnel, but it should not be presented as a durable multi-instance deployment.

Accomplishments that we're proud of

  • One working path from landlord source upload to confirmed publication, current Web and Mobile use, customer-package preparation, and landlord weekly reporting.
  • Source-category handling that preserves official building facts when a document is only suitable as reference or manual review material.
  • A shared current-information contract across Admin, common Web workflows, and Mobile.
  • Explicit version and provenance controls that prevent stale plan reuse.
  • Human approval and configured-recipient gates for every external-facing output.
  • A synthetic sample that a judge can download and run without credentials.
  • 289 passing automated tests, five strict workspace typechecks, and a passing repository package validation.
  • A reusable thin skill that captures how a domain expert and Codex moved from conversation to tested workflow.

What we learned

The most useful role for AI in a sensitive workflow is not autonomous authority. It is the reduction of reconstruction work.

Codex can help a person surface hidden rules that are obvious inside the job but absent from a software brief: which document is current, who can make it official, which information may leave the company, what belongs in a landlord meeting, and what a mobile user must see without scrolling or decoding system language.

We also learned that conversation becomes valuable only when it is closed by evidence. A correction should become a product rule, a working path, and a test that the real user can challenge again.

What's next

The immediate next step is a limited pilot with the real leasing user using safe test data. His feedback will determine which steps still feel artificial, which building fields are missing, and whether the weekly landlord report matches the meeting itself.

After that, LeaseFlow needs durable multi-user persistence, enterprise identity and role provisioning, an approved document repository, and controlled Microsoft 365 integration. Those production integrations are not implemented or claimed here.

Codex for Everyone can then be applied to a second domain. The goal is to test whether the same thin loop—evidence, workflow reconstruction, authority mapping, one vertical slice, real-user correction, and verification—can remove a different person's recurring bottleneck without requiring that person to become a software product manager first.

Built With

Share this project:

Updates