Updated: 2026-10-11. Owner: Jay Liu. Audience: senior enterprise/full-stack and applied AI hiring teams.
The opening description presents enterprise platforms and applied AI together, backed by 18+ years in enterprise software and ERP/CRM. C#/.NET, React/TypeScript and Python support current agent orchestration, engineering-delivery workflows, knowledge retrieval and speech-processing work. Earlier financial, retail and hospital-system experience remains explicit in the career record.
ERP/CRM and retail business-system experience are explicit in the owner's original career record. The retail role used .NET Core, REST APIs, MySQL and Vue.js. Earlier leadership roles covered enterprise middleware, database architecture, server delivery and hospital management software. The 2022 and owner-reviewed 2026 career records supply these facts. No scale, revenue, performance or customer-result figures are added. The established 18+ years search identity remains. The public CV has been replaced with the evidence-aligned two-page version described in recruiting evidence.
The hero introduces engineering value without employer emphasis. About and the career record distinguish FNZ delivery work from independent AI projects. General delivery responsibilities use the owner’s clarification, existing CV and official Analyst Developer role context; confidential assignments and internal implementation details are excluded. The site contains no personal-business promotion or company-service links. Third-party papers are not displayed as portfolio evidence.
Four text-led engineering cases appear before career experience on the homepage. Each identifies the business problem, Jay’s role, a constraint/decision, current stage and inspectable evidence. The full thirty-example gallery is retained on /projects/ and /zh/projects/. Desktop has three columns, tablet two and phone one. Each card contains a small preview, a title, a brief purpose statement and a public website or GitHub link where appropriate. The displayed count is the selected inventory, not a claim about the total number of projects completed.
The gallery begins with a conversational-device backend, Web Builder Agent and Idea Builder. Agent workbenches, writing assessment, reusable tooling and voice workflows follow, then the retained operations ERP, transactional CRM, industry websites and historical projects. The ERP description is backed by its actual planning/catalogue/order/procurement/reminder interfaces. It is labelled a development instance; no claim about production acceptance or commercial scale is made.
The financial-project domain serves a holding page; its screenshot is historical and the link says “Current site”. Other historical destinations had no usable secure public page at the last check and have no links. Public customer-domain links are owner-authorised. All client labels remain anonymous industry descriptions. Private application URLs, source repositories, local paths, real customer records and credentials are excluded.
Eleven screenshots are intrinsic 320×200 WebP files, rendered at no more than 320 CSS pixels. They have no clickable image wrapper, lightbox, full-size sibling or high-resolution srcset. Resolution and redaction provide the privacy boundary.
Existing exports mask brands, contacts and private session/user areas. The historical medical portrait is blurred and its signature covered. The CRM preview is an empty-data demo; the memory preview is CLI help without personal records. The new ERP preview renders the actual dashboard and shell components with an anonymous account, empty collections and zero counters. It makes no API calls and accesses no database. Nineteen generated covers are screenshot-style interface mockups with accurate generated-image alternative text. Separate per-image caption lines remain absent. The gallery introduction describes the mixed media, and generated previews have a compact in-frame “AI interface illustration” label. They use generic demo content to explain software purposes, without reproducing private interfaces or implying actual screenshots.
Original captures, temporary render scripts, application HTML and proof files remain outside Git/public output. Only reviewed low-resolution image exports are published. Names and alternative text use anonymous project IDs and industry descriptions. Any replacement requires a fresh privacy review.
English and Chinese expose the same career facts, 30 projects, 11 screenshots, 19 labelled interface mockups, status and links. Project descriptions stay short; detailed technical lists are omitted from the cards. All content remains in initial HTML, including when JavaScript or images are disabled.
The established homepage title, canonical and alternate URLs, schema identity, old anchors and unchanged CV URL are preserved. Both homepage descriptions are deliberately rewritten to distinguish employment from independent AI work. Project indexes have their own reciprocal locale metadata. No claims about hiring conversion, search rankings or field performance follow from a successful build or local lab result.
The additional sixteen cases were selected from original code, README files, commit authorship and current system modules; folders containing only third-party source or undeveloped plans were excluded. A single multi-package SDK counts as one case. Descriptions report capabilities rather than unverified production scale, outcomes or model-training work.
The public HiLily website identifies Jay as a volunteer AI Coach for architecture, memory and audio guidance. Both the featured case and complete gallery use that team role, name the functional MVP / pilot-preparation stage and link to demonstrations and the team. No sole-authorship, health-outcome or latency claim is used. Current sources and content decisions are recorded in positioning.md.
The Web Builder workflow was checked against the running source: research/build transitions, reusable stages, evidence evaluation, retries and review gates. Agent profiles recognise Claude Code, Codex and OpenCode. The portfolio describes workflow/profile compatibility; it does not assert that all engines have been executed through the current runner. Private runtime endpoints and source repositories are omitted.
Voice2Task, live interpretation, AirMicLAN and local narration describe capture, segmentation, transcription, translation, LAN streaming and delivery integrations. VoxCPM2 and OmniVoice remain attributed as existing model/engine integrations.
CodexView and AgentMeshKit source links were checked as public, with Jay's commit authorship verified. Other new private systems include guided multi-agent delivery, writing assessment, agent configuration sync, persistent workspaces, browser testing, ESP32 HID control, reviewed editorial workflows and Android media delivery. Their interface mockups use generic functionality without customer records or private interfaces.
The existing small-image restrictions remain in force. A separate text and interactive walkthrough uses fictional notes in a browser-only simulation, with no production connection. It illustrates implemented retry and revision rules and explains internal usage and human review. All six languages have equivalent content and controls. This explicitly labelled simulation is additional evidence presentation; it does not claim to be an actual production session.