• Deterministic DAG Scheduling & Deadlock Prevention: Client-side Kahn's topological sorting algorithm with cyclic dependency detection and dynamic branch pruning.
• Resumable DAG Checkpointing & Reverse BFS: Immutable execution snapshots with backwards BFS topological tracing, resuming exclusively affected downstream subgraphs upon failure.
• Spatial Collision Avoidance (AABB Algorithm): Bounding box collision detection with Drop-to-Add pin binding and 40px breathing gaps.
• Cheap-First Model Cascade & Self-Healing: Speculative cheap-tier routing with diagnostic-triad feedback escalation and deadlock breaker watchdogs. Measured on a 30-case A/B: schema compliance 90.0% → 100.0%, escalation rate 0% → 10%, token overhead +21.8% (paired McNemar p = 0.25 — not statistically significant; n = 30 with a 3-case treated cohort). Reported as a directional architectural finding, not a proven effect.
• Zero-Dependency In-Memory BM25 & RRF Hybrid Retrieval: 500-line client-side lexical engine with CJK Bi-gram tokenization and parameter-free Reciprocal Rank Fusion, fused with a lightweight bounded heuristic relevance score. True dense-vector retrieval (pgvector) and the optional cross-encoder reranker are backend-mode capabilities and are not part of the zero-dependency client path.
• Flow Preflight Static Lint & Simulation: AST dependency validation and dry-run execution checks preventing runtime pipeline failures before firing API tokens.
• Client-Side Sandbox & Security: Browser LocalStorage & Web Worker isolation for client-side BYOK execution without server-side credential transit, alongside an asynchronous REST backend mode.

• Role: Product Engineer / Product Architect
• Focus Areas: AI Workflow Orchestration, DAG Execution Engines, Distributed SaaS Architecture, Production RAG Systems
• Core Engineering Principles: Deterministic Execution, High Reliability, Local-First Privacy, Zero-Server Credential Transit (BYOK)

Having operated across the full lifecycle—from massive-scale concurrent interactive platforms and global growth engines to developer infrastructure and open-source commercialization—I firmly believe that grand product abstractions must ultimately converge into robust, working code and practical user experiences.

In this AI era, I have intentionally returned to the front lines as a hands-on builder, embracing AI Pair Programming and "Learning by Doing":

• Beyond Chat Wrappers: Moving past simple prompt boxes toward node-graph topological execution, decoupled state slices, and local agent primitives.
• Model vs. Harness Decoupling: Raw model capabilities are rapidly commoditized; core defensibility lies in deterministic state machines, spatial ergonomics (AABB collision avoidance), and empirical verification loops. (See: 《Resisting Mode Gravity》 · 《Zero-Dependency BM25 & RRF Hybrid Search》)
• Milestones & Extreme Testing: Grounding architecture in real workflow trade-offs, discovering underlying system boundaries and deadlock prevention through bidirectional AI collaboration, and personally verifying edge cases under extreme conditions.
• Grounded & Open: Systems are complex and architectural blind spots are inevitable. Always grounded in humility, and warmly welcoming peer discussions, architectural critiques, and code reviews.

After leaving JiHu GitLab, I channeled over a decade of engineering, product, and growth experience into multi-platform commercial operations and our independent R&D studio.

Initially, operations relied heavily on manual effort. As order volume and customer segments scaled, off-the-shelf platform tools hit a wall. I brought in AI Workflow Orchestration and AI Pair Programming: first building automated mini-programs for acquisition and community engagement, then tackling the hardest bottleneck—unifying fragmented orders, logistics, and customer identities across heterogeneous channels.

Years ago, cross-channel ETL and unified membership mapping would have required a dedicated R&D squad. Pairing with Claude, I evolved the stack step by step: starting with n8n script chains, migrating to Dify, and after hitting state-control limits in generic tools, architecting my own local visual orchestrator powered by a deterministic DAG Engine and State Machine.

Throughout this milestone co-discovery process, I defined the product ergonomics and operational scenarios (multi-flow drawer isolation, zero-config local LocalStorage privacy execution), while AI helped surface deep algorithmic edge cases (Kahn's topological cycle detection, React Flow 60fps zero-dirty-rerender state slicing, and AABB spatial collision avoidance). After stripping and anonymizing proprietary business logic, I open-sourced the core engine as PatchCat.

With the underlying automation pipelines running reliably in production, my focus remains dedicated to advancing deterministic AI workflow orchestration, client-side DAG state machines, and next-generation AI-Native product architectures.

JiHu GitLab was my first all-remote, international open-source company—and still one of my favorite engineering cultures. Stepping in to lead Ops product lines and China SaaS, I was immersed in GitLab's radical "Dogfooding" ethos, where distributed global teams collaborate asynchronously through transparent handbooks, Issues, and Merge Requests.

To unlock China SaaS growth, I combined platform product engineering with full-funnel GTM mechanics: managing both the official marketing site and developer community, scaling organic acquisition via technical SEO, and overhauling the trial onboarding funnel. Because GitLab's architecture couldn't rely on off-the-shelf third-party trackers, I built an independent telemetry and full-funnel conversion analytics system from scratch, while initiating early enterprise AI DevSecOps explorations.

Later, amid macro market shifts, the company consolidated resources toward its core self-managed foundation, strategically streamlining local exploratory lines. That experience taught me to evaluate every product line not just by local feature execution, but through the macro lens of corporate cash flow, unit economics, and strategic gravity.

This chapter was a bold leap into deep-tech hardware: building high-precision pharmaceutical spectroscopy instruments and companion workstation software for research labs. Backed by early commercial intent orders, I dove headfirst from internet software into bare-metal industrial R&D.

In 2019—long before vibe coding—mastering the full hardware-software stack meant starting at STM32 registers and stepper motor drivers: reading datasheets, analyzing timing diagrams, and probing signal waveforms with an oscilloscope whenever a motor stalled. From the bare metal up, I worked across custom real-time scheduling, serial/bus communication protocols, and the C# .NET desktop spectroscopy workstation and lab data platform.

When external supply-chain disruptions hit, custom mechanical components were delayed for months. I took the dev boards and optical modules home to keep debugging firmware and standardizing our Jira R&D workflows. While hardware supply cycles eventually slowed commercial scaling, that period forged an irreversible confidence: no technical black box is too intimidating to learn from scratch.

After Bilibili, I wanted to step outside pure consumer traffic plays and test myself in B2B2C monetization—joining the innovation arm at Alo7 (Qinwen Network).

We built a full-stack growth and retention suite sold directly to educational institutions: WeChat Mini-Programs driving daily reading check-ins and viral community loops on the front end, backed by a custom-built lightweight CRM, user profiling engine, and licensed video curriculum. Systematic A/B testing lifted registration conversion by over 30%. Around the same period, I also explored offline QR-terminal traffic monetization integrated with major e-commerce APIs (Alibaba/JD), learning firsthand why high-urgency offline micro-moments clash with slow-decision e-commerce rebate funnels.

As regulatory headwinds hit the education sector, the innovation line was scaled back. Stepping outside the traffic greenhouse of mega-platforms taught me a sobering truth: when you no longer have guaranteed organic traffic, matching user psychological intent and macro industry cycles matters far more than feature polish alone.

Joining Bilibili during its hyper-growth phase, I brought my mobile attribution experience from Ctrip to overhaul Bilibili's entire mobile web (H5) platform.

My goal was simple: make it industry-leading. During the architectural rewrite, I proactively embedded channel telemetry and distribution package attribution ahead of business requests. When the marketing team later came asking for campaign tracking, they were stunned to find the entire attribution infrastructure already live. Having end-to-end observability also enabled us to detect abnormal traffic patterns and block unauthorized third-party apps hotlinking media streams. Once observability is in place, nothing stays hidden.

Yet that excellence created a strategic paradox: strategically, Mobile Web existed as a top-of-funnel acquisition bridge to drive native App installs. The smoother and more complete I made the H5 playback experience, the less incentive users had to download the native App.

It was a defining lesson in product strategy: local craftsmanship always yields to company-level strategic gravity—though proving we could engineer a top-tier high-concurrency mobile web stack remains a proud milestone.

During the post-delivery maintenance phase of the bank core project, a late-night technical forum exchange turned into an interview invite—which turned out to be Ctrip. I joined initially on a hybrid PHP Engineer / Marketing PM contract before transitioning full-time into product management. Having written production code myself meant engineering teams could never bluff me on API contracts or sprint estimates.

Right as James Liang returned to lead Ctrip's mobile-first resurgence, I was handed a foundational challenge: standardizing event telemetry and order channel attribution across 4 fiercely independent Business Units (BUs). After aligning the cross-BU data foundation, I spearheaded two ahead-of-their-time growth engines: a site-wide dynamic H5 engagement system, and seamless H5-to-App invocation preserving full channel context—an early production implementation of Deferred Deep Linking that lifted order conversion across the funnel by 70%~80%.

After taking those 0-to-1 growth infrastructures from conception to stable production, an invitation from hyper-growth Bilibili led me to my next challenge in massive-scale content distribution.

My first job after graduation in Shanghai started with an unexpected interview discussion on Linux 0.11 kernel memory management—proving I actually read source code rather than memorizing textbook answers landed me an immediate offer.

Working on a major commercial bank's Data Centralization Platform (IBM DB2 mainframes, C, PL/SQL, and banking middleware) taught me lifelong reverence for the word "Production." During a critical cutover window, traffic switchover triggered queue congestion alerts as downstream responses neared timeout thresholds. Under strict rollback SLA pressure, our team rapidly traced the bottleneck to an unadjusted queue-width parameter in the high-concurrency config. Dynamically tuning the queue parameter and watching the backlog drain smoothly within the cutover window was an unforgettable baptism by fire.

Months of 2 a.m. data center shifts—monitoring green-on-black terminal crontab batch pipelines and end-of-day reconciliation jobs—instilled a permanent instinct for high availability and strict data consistency before I leaped into the rising mobile internet wave.