Jay Liu.Senior Software Engineer

Customer-service teams need workflows shaped around their own follow-up, analysis and approval problems. A useful CRM must connect business context with the next action.

My role

I designed and built the CRM, API and CLI as a customisable foundation for AI-assisted service workflows, linking customer history, projects, invoices and versioned notes.

Constraint & decision

My integration approach is to connect reminders and business analysis to proposed actions, with human approval before consequential execution. Stable request keys, version checks and attributed updates make agent writes traceable and protect shared records.

Current evidence

The internal system includes follow-up reminders, shared business records and controlled API/CLI writes. Broader AI analysis and decision workflows are extension areas tailored to each business. The fictional-data demo shows the implemented write-safety rules.

Receipts and bank statements arrive in different formats. Preparing a reviewable ledger requires matching entries without losing the original evidence.

My role

I built the reusable skill and supporting tooling for evidence capture, ledger preparation, reconciliation and export.

Constraint & decision

Unclear matches must remain visible. Preserve source files, normalise records into JSON and leave ambiguous items for human confirmation before preparing exports.

Current evidence

Public workflow instructions and scripts cover Xero CSV exports and accountant review packs. Financial records and final decisions remain subject to human review.

Hardware development and extending it into a complete software system are central challenges for this voice companion, especially echo, noisy environments and conversational response time.

My role

As a volunteer AI Coach, I contribute technical guidance on the hardware-to-software AI pipeline, microphone audio processing and low-latency speech-to-speech model integration. Hardware and product delivery are team efforts.

Constraint & decision

The work includes echo cancellation with single- and dual-microphone setups, recognition in noisy environments, and audio-in/audio-out language models. I help the team work through these connected constraints across device audio, model integration and system extension.

Current evidence

The project website includes demonstrations and my team role. The team reports a working MVP with multiple iterations completed and is preparing for real-world pilots.

Coding agents produce different transcript formats. Reviewing tool calls and nested sessions should not require a different viewer for every agent.

My role

I developed the transcript adapters, React components and Markdown CLI as reusable packages.

Constraint & decision

Keep format parsing separate from presentation. Adapters produce a shared event model that supports both a visual transcript and compact text, preserving tool and subagent context.

Current evidence

The public repository contains the three packages, usage examples and test fixtures. The implementation can be inspected independently of private workbench projects.