KAYN / THE JOURNEYIDEAS INTO REALITY

Always building.

Products I've built. Teams I've grown with. Code I've contributed.

01 / PROJECTS

Built from an idea.

Explore the work below

Product Proposal / System Architecture Design

Holonic Asset (2D Pixel Game Asset Generation & Management Platform)

July 2026 ~ Present

A pixel-art 2D game asset generation and management platform for indie developers and small teams. Built around Project, Asset, Record, and Generation domain models, establishing an end-to-end workflow from project-level AI context, async generation, version rollback to engine export (Unity/Godot).

  • React
  • TypeScript
  • Tailwind CSS
  • Go
  • Echo
  • PostgreSQL
  • Oxlint
  • MCP
View codeWatch demo

02 / EXPERIENCE

Better, together.

Learning through real work
2026July 2026 ~ Present

Qiniu

AI Infrastructure Intern

  • LaaS Application Server: Contributed to Qiniu LAS Agent sandbox and service interfaces, optimizing instance lifecycle reliability as well as startup and build performance

  • MaaS Platform Engineering & Orchestration: Collaborated with the core team to develop the Sufy MaaS Platform serving 10,000+ active users, taking charge of model serving pipelines, async task orchestration, and high-availability API routing to ensure resilience and low latency

The full story

Qiniu is a leading cloud infrastructure, AI, and scenario-based audio/video APaaS provider in China. During my internship as an AI Infrastructure Intern, I contributed to Qiniu LAS (LaaS Application Server) by supporting Agent sandbox services and optimizing instance lifecycle reliability, startup, and build performance, while collaborating with the core team to build the Sufy MaaS Platform serving 10,000+ active users across model serving pipelines, async task orchestration, and high-availability API routing. Additionally, I served as Team Captain to spearhead the architecture design, domain modeling, and 4-tier engineering quality loop for a 2D pixel game asset generation platform (Holonic Asset).

  • Product Architecture & Domain Modeling: Led architectural design of the 2D game asset AI platform; established Project creative kernel, heterogeneous Assets, async Task orchestration, and immutable Record snapshot models to eliminate style drift
  • System Design & Modular Monolith: Adopted a modular monolith architecture decoupling Workspace, Generator, Processor, and Storage domains with shared transactional boundaries to ensure high cohesion and fast iteration
  • Engineering Quality Loop: Designed and enforced an end-to-end quality loop (structured Issue specs, Lefthook pre-commit checks, GitHub Actions CI, and backend-driven contract generation) to guarantee robust delivery
  • Team Leadership & Coordination: Served as Team Captain, leading requirement decomposition, task scheduling, architecture reviews, and Code Reviews to ensure timely, high-quality deliverables
2025August 20, 2025 ~ December 31, 2025

Neka

Full Stack Intern

  • Worked across the full stack of AI product prototypes, covering frontend interaction design, backend APIs, database modeling, and end-to-end integration

  • Designed and built conversational workflows and multi-role testing capabilities to support state transitions, action triggering, and result validation in complex scenarios

The full story

Neka is a modular digital art creation and trading platform. The company is committed to enabling all users to customize art assets with no barrier to entry, dramatically increasing productivity in this field, reshaping traditional transaction models, and becoming a global next-generation content production platform and super marketplace. In less than a year after launch, the platform reached 1.2 million registered users, accumulated tens of millions of visits, and attracted users from more than 200 countries and regions worldwide.

  • Built internal configuration and debugging back-office tools for visual management of core rules, states, entities, and workflow logic
  • Integrated external platforms and internal business systems to enable conversation-driven operations, task flow execution, and data synchronization
  • Implemented persistence for conversation context, history, configuration, and credentials to improve reliability, traceability, and maintainability
  • Independently delivered full-stack capabilities for a conversational AI project management assistant, including message intake, workflow orchestration, data storage, and an admin console
2025June 19, 2025 ~ August 19, 2025

Demoway

Full-Stack Development Intern

  • Maintained and optimized the core Html to Figma conversion engine, diving deep into the end-to-end AST pipeline from DOM traversal and Computed Style extraction to Figma nodes and AutoLayout mapping

  • Delivered full-stack enhancements for a V0-like web generation system, designing real-time code rendering components and backend data synchronization channels

The full story

Demoway focuses on AI+Design SaaS solutions, serving leading global enterprises (Alibaba, Tencent, ByteDance, Amazon, Microsoft, OpenAI, Google) as well as creative studios. Its core offerings include Html to Figma, Web to AI, Screenshot to design, and Copy to Design, converting existing web assets into design systems across Figma, MasterGo, and Pixso to significantly accelerate design workflows. During my internship as a Full-Stack Development Intern, I actively participated in maintaining and optimizing the core Html to Figma product, mastering the end-to-end AST data transformation pipeline from DOM tree traversal and Computed Style extraction to Figma AutoLayout. In addition, I delivered full-stack capabilities for an open-source V0-like AI webpage generation system and integrated upstream/downstream services via SDKs.

  • Integrated LLMs and refined prompt engineering workflows to improve the accuracy and structural robustness of generated code
  • Designed and encapsulated cross-system SDKs to bridge generated code into downstream design toolchains (Figma, MasterGo, etc.)
  • Participated in the full-stack architecture design and end-to-end delivery of the AI+Design SaaS platform

03 / OPEN SOURCE

A little code. A shared world.

23Repositories
40Pull requests
7Issues

terax-ai

Fixed terminal focus restoration after dismissing Terax AI's search panel for a smoother keyboard-driven search flow.

1 PRN/A
Explore contributions

Fixed terminal focus restoration after dismissing Terax AI's search panel for a smoother keyboard-driven search flow.

curl

Fixed libcurl multi_socket API waiting indefinitely when a server stays silent by scheduling the response deadline in the event loop; added FTP regression coverage.

1 PR1 Issue1 CommitN/A
Explore contributions

Fixed libcurl multi_socket API waiting indefinitely when a server stays silent by scheduling the response deadline in the event loop; added FTP regression coverage.

react

Fixed React's stable useState setter dependency handling in memo preservation and proposed preserving DevTools profiling across soft navigations.

2 PRN/A
Explore contributions

Fixed React's stable useState setter dependency handling in memo preservation and proposed preserving DevTools profiling across soft navigations.

gin

Added AIP custom verb path support to Gin's router for matching custom HTTP verb routes.

1 PRN/A
Explore contributions

Added AIP custom verb path support to Gin's router for matching custom HTTP verb routes.

kratos

Fixed Protobuf field alias resolution in Kratos configuration scanning, ensuring snake_case and camelCase field aliases unmarshal properly into Protobuf messages.

1 PRN/A
Explore contributions

Fixed Protobuf field alias resolution in Kratos configuration scanning, ensuring snake_case and camelCase field aliases unmarshal properly into Protobuf messages.

godot

Fixed duplicate Button presses from touch-emulated mouse input and improved inheritance of untyped GDScript export override return types.

2 PRN/A
Explore contributions

Fixed duplicate Button presses from touch-emulated mouse input and improved inheritance of untyped GDScript export override return types.

gemini-cli

Restored Gemini CLI resume support for legacy sessions by improving session discovery and timestamp fallbacks, with tighter explicit-resume error handling.

1 PRN/A
Explore contributions

Restored Gemini CLI resume support for legacy sessions by improving session discovery and timestamp fallbacks, with tighter explicit-resume error handling.

gogen

Added static value member resolution and import support to the Go+ gogen compiler, completing the downstream compilation path for XGo static values.

2 PR1 IssueN/A
Explore contributions

Added static value member resolution and import support to the Go+ gogen compiler, completing the downstream compilation path for XGo static values.

xgo

Extended the XGo compiler with static value declaration syntax, including T.name and classfile .name shorthand, and exposed HasStatic for downstream AST lowering.

2 PRN/A
Explore contributions

Extended the XGo compiler with static value declaration syntax, including T.name and classfile .name shorthand, and exposed HasStatic for downstream AST lowering.

controller-runtime

Added indexed cache reads to controller-runtime informertest, expanding test coverage for indexed cache queries.

1 PRN/A
Explore contributions

Added indexed cache reads to controller-runtime informertest, expanding test coverage for indexed cache queries.

kueue

Optimized Workload Assignment (WAS) planner failure reporting in Kubernetes Kueue by batching PodSet assignment and flavor attempt updates, reducing redundant scan overhead.

1 PRN/A
Explore contributions

Optimized Workload Assignment (WAS) planner failure reporting in Kubernetes Kueue by batching PodSet assignment and flavor attempt updates, reducing redundant scan overhead.

kubernetes

Refactored Kubernetes CSI VolumeAll tests to prevent goroutine leaks with test-owned stop channels, and optimized client-go reflector initialization by using any resource version for initial WatchList requests to prevent stalls against lagging API server caches.

2 PRN/A
Explore contributions

Refactored Kubernetes CSI VolumeAll tests to prevent goroutine leaks with test-owned stop channels, and optimized client-go reflector initialization by using any resource version for initial WatchList requests to prevent stalls against lagging API server caches.

dify

Contributed stability improvements to Dify: modernized integration-test log waits, fixed the RAG DSL import transaction issue, and pinned the web Docker Node version.

3 PRN/A
Explore contributions

Contributed stability improvements to Dify: modernized integration-test log waits, fixed the RAG DSL import transaction issue, and pinned the web Docker Node version.

plane-mcp-server

Added the list_project_issues tool to Plane's MCP subproject, extending project management integration coverage.

1 PRN/A
Explore contributions

Added the list_project_issues tool to Plane's MCP subproject, extending project management integration coverage.

milvus

Authored core query-engine and scalar index optimizations in Milvus: In PR #53960, accelerated regex filtering on sealed VARCHAR by extracting bounded Boolean literal conditions from RE2, performing per-segment cost-based selection via FM-Index count, and validating candidates with canonical RE2 recheck, achieving up to 44x speedup on selective patterns with provable soundness and fallback safety; in PR #53901, optimized sorted scalar-index IN / NOT IN predicates using sorted batch lookup with cursor-driven galloping search across 8 scalar types, achieving up to 33.74x speedup.

2 PR2 IssueN/A
Explore contributions

Authored core query-engine and scalar index optimizations in Milvus: In PR #53960, accelerated regex filtering on sealed VARCHAR by extracting bounded Boolean literal conditions from RE2, performing per-segment cost-based selection via FM-Index count, and validating candidates with canonical RE2 recheck, achieving up to 44x speedup on selective patterns with provable soundness and fallback safety; in PR #53901, optimized sorted scalar-index IN / NOT IN predicates using sorted batch lookup with cursor-driven galloping search across 8 scalar types, achieving up to 33.74x speedup.

moby

Improved Moby iptables rule management and Swarm API documentation by fixing a chain-head race and documenting the Availability field.

2 PRN/A
Explore contributions

Improved Moby iptables rule management and Swarm API documentation by fixing a chain-head race and documenting the Availability field.

oxc

Contributed to Oxc's linter, transformer, and minifier: improved React component export rules and static-block lowering semantics, simplified single-property object destructuring in the minifier, and added regression and conformance coverage.

3 PRN/A
Explore contributions

Contributed to Oxc's linter, transformer, and minifier: improved React component export rules and static-block lowering semantics, simplified single-property object destructuring in the minifier, and added regression and conformance coverage.

redis

Contributed command validation and bitop performance improvements to Redis: in PR #15957, extended the request/response validator to match accepted requests against argument declarations; in PR #15920, restructured BITOP execution paths to remove shared bytewise tail loop overhead and resolve performance regression for short unaligned values.

2 PR1 IssueN/A
Explore contributions

Contributed command validation and bitop performance improvements to Redis: in PR #15957, extended the request/response validator to match accepted requests against argument declarations; in PR #15920, restructured BITOP execution paths to remove shared bytewise tail loop overhead and resolve performance regression for short unaligned values.

Easydict

Implemented native MDict (.mdx/.mdd) import, lookup, and WebView rendering without format conversion. Used binary search, on-demand decompression, LRU caching, and threshold-pruned Levenshtein matching, supported absolute audio paths for system dictionaries, and continued fixing linked entries, rendering interactions, anchor navigation, and dark mode issues.

6 PR1 IssueN/A
Explore contributions

Implemented native MDict (.mdx/.mdd) import, lookup, and WebView rendering without format conversion. Used binary search, on-demand decompression, LRU caching, and threshold-pruned Levenshtein matching, supported absolute audio paths for system dictionaries, and continued fixing linked entries, rendering interactions, anchor navigation, and dark mode issues.

tokio

Bound sharded blocking-queue batches in the Tokio runtime to improve spawn_blocking fairness under high concurrency.

1 PRN/A
Explore contributions

Bound sharded blocking-queue batches in the Tokio runtime to improve spawn_blocking fairness under high concurrency.

next.js

Fixed Next.js custom server basePath configuration not reaching the internal render server and causing 404 responses.

1 PRN/A
Explore contributions

Fixed Next.js custom server basePath configuration not reaching the internal render server and causing 404 responses.

zed

Corrected the Zed Agent tool classification and UI icon for create-directory so its mutating filesystem behavior is represented accurately.

1 PRN/A
Explore contributions

Corrected the Zed Agent tool classification and UI icon for create-directory so its mutating filesystem behavior is represented accurately.

milvus-backup

Fixed secondary cluster restore synchronization in Milvus Backup: in PR #1289, added target-state guard checks before CDC stream initialization, preventing checkpoint divergence and unrecoverable transaction blocking during backup recovery.

1 PR1 IssueN/A
Explore contributions

Fixed secondary cluster restore synchronization in Milvus Backup: in PR #1289, added target-state guard checks before CDC stream initialization, preventing checkpoint divergence and unrecoverable transaction blocking during backup recovery.

04 / PUBLICATIONS

Research, on the record.

Design and Implementation of an Intelligent Bookkeeping System with OCR and LLMs

Abstract

Background Modern personal bookkeeping still suffers from low conversion rates from free-form receipts to usable ledgers. Objective We build a cross-platform system that turns unstructured inputs into validated ledgers with minimal user effort. Method Smart Accounting combines Optical Character Recognition (OCR), Vision-Language Models (VLM), and Large Language Models (LLM) with constrained JavaScript Object Notation (JSON) extraction: the prompt forces a fixed key set (amount, currency, date, category, vendor, paymentMethod, memo, source), enum-typed fields, and three retry slots; backend validation replays the error message to the model and rejects non-conforming payloads. Results On a 180-receipt internal benchmark (CORD and DocVQA subsets), the hybrid selector improved line-item F1 to 92.1% versus 89.4% (OCR+LLM) and 86.7% (VLM only), while reducing schema violation rate to 0.6%. Latency remained under 2.1 s median on desktop and 2.9 s on mobile. Significance The system delivers interpretable, schema-safe financial drafts, streaming narrative reports, and deployment portability (desktop, web, mobile), offering a practical blueprint for intelligent personal finance applications.

Index Terms

OCR, Vision-Language Models, LLM, Intelligent Accounting, Personal Finance, Cross-Platform Systems

KAYN / STILL EXPLORING