Project Ledger · AI Project Progress Assistant

Turn spoken updates and screenshots into reviewable, traceable project records

多模态理解Codex CLIMCPTypeScriptReactCloudflare WorkersD1幂等写入

211/211

Offline regression tests passed at delivery

MCP tool interfaces preserved

Calendar views: daily intake and project timeline

Project Summary

Designed and shipped an owner-private project ledger for managing progress across multiple projects. It receives assistant-transcribed updates through dot and MCP, combines task dependencies with daily intake and project timeline views, and turns image content into editable previews. Updates require confirmation and an independent readback. A real reservation screenshot test correctly identified an existing event and produced a duplicate-only preview without creating another record.

Background & Challenge

Project updates were scattered across conversations, screenshots and files, with no single reliable record of task state. Dependencies could be overlooked, and upload dates could be confused with event dates. Understanding an update with AI was only one part of the problem: the product also needed to show whether a change was saved, prevent repeated writes and recover safely from failures.

My Role

Led product definition, workflow design and acceptance criteria, using Codex to implement and validate iterations. Distinguished execution tasks from scheduled events and reference material. Established a preview, confirmation, save and independent-readback workflow, while preserving existing data and integration contracts. Prioritized trustworthy state and dependency handling before expanding calendar and image-understanding capabilities.

AI Integration

Built with a TypeScript/React frontend, a Worker service and a private D1 database. Seven MCP tools connect dot to reading, previewing, committing, receipts and undo. A local bridge invokes the official Codex CLI with the existing ChatGPT login, using isolated recognition tasks and constrained structured output. The server validates fields, evidence, request identifiers and existing records before an explicitly confirmed transaction. Idempotency, version checks, retained drafts and independent readback support consistent updates.

Implementation Process

Phase 1: Define a unified model for projects, tasks, events and reference material

Phase 2: Implement previews, confirmation, transactional saves, receipts and independent readback

Phase 3: Connect dot through MCP while preserving existing interfaces and idempotent retries

Phase 4: Add prerequisite navigation, completion-based unlocking and state constraints

Phase 5: Release daily intake and project timeline views with automatic task nodes

Phase 6: Integrate local Codex image understanding, test reservation recognition and duplicate previews, then update the existing private site

Key Deliverables

Deployed private project-ledger web application

Existing dot/MCP integration with seven tools preserved

Task dependency navigation and prerequisite locking

Daily intake calendar and project timeline

Image-understanding module reusing the local Codex login

Editable previews, existing-record duplicate checks and idempotent saves

211 offline regression tests, acceptance evidence and project handoff documentation

Reflection

Recognition success is different from a confirmed save, and a successful reservation does not mean an event has already happened. The interface needs to distinguish AI interpretation, stored records and real-world facts. Reusing the local Codex login reduced repeated authorization, but still requires an online computer, local bridge and browser pairing. The real-image acceptance test validates one reservation and duplicate-preview scenario; broader reliability needs a larger set of conflicting dates, completion evidence and cross-project examples.

Interested in this technology?

Whether you are a potential partner, recruiter, or technical evaluator — start a conversation anytime.

Get in Touch