w.winhok
中文

winhok.

AI Agent & Full-Stack Engineer

I build production AI products, reusable agent infrastructure, multi-agent orchestration and governed retrieval systems.

Watch product walkthroughs View Engineering Work

If you found me through Upwork, please keep pre-contract communication on Upwork.

Finahead financial decision walkthrough

Four connected capabilities

Production products, an open-source runtime, orchestration and retrieval.

Knolume

Multi-Product AI Platform

A private multi-product AI platform powering Interview and Finahead, with shared account access and agent execution while keeping product data and workflows isolated.

PRODUCTION DEPLOYED · Interview & FinaheadVIDEO · Two product walkthroughsARCHITECTURE · Shown here

Interview and Finahead are two deployed products that share account access and agent execution while retaining their own user journeys and data boundaries.

Two products, distinct responsibilities.

The shared platform provides account access and agent execution. Interview and Finahead keep their own user tasks, data and rules.

Interview

Production deployed

Walkthrough uses demo candidate data.

Practice grounded in a candidate’s own material: résumé-based questions, follow-up rounds, saved progress, reports and focused practice.

Interview keeps its own session progress and candidate material while using the shared account and execution boundaries.

25-second walkthrough: résumé-based practice, a follow-up question and actionable feedback.
Walkthrough in text

A candidate resumes a backend reliability practice round, answers with server-side ownership checks, then receives a follow-up asking for the exact tests behind that claim.

Finahead

Production deployed

Walkthrough uses synthetic financial data.

A personal finance workspace connecting transactions, forecasts, goals and scenario comparisons. The ledger’s versioned calculations remain separate from optional AI assistance.

The product is integrated with shared account access while retaining its own ledger ownership and command rules. A suggested change still needs the product’s permission and version checks.

29-second walkthrough: financial overview, purchase impact and assistant guidance.
Walkthrough in text

The ledger compares an ¥8,000 laptop purchase with cash, investable funds and a travel goal. The assistant also flags a longer-horizon cash shortfall instead of recommending the purchase.

Knolume Runtime

Composable execution primitives for AI applications

A public TypeScript library for agent runs, tools, context, memory, retrieval, guardrails and sub-agents. Host applications provide models, tools, identity and product policy.

ARCHITECTURE · Shown hereSOURCE EVIDENCE · Reviewed a19f728
  • Agent loopsrc/harness/agent/loop.ts

    Run events, model steps, cancellation and tool-runtime composition.

  • Tool execution boundarysrc/tools/execution-pipeline.ts

    Argument validation, permissions, approvals and outcomes.

  • Guardrail runnersrc/harness/guardrails/runner.ts

    Buffered checks, effect gating and recovery handling.

Public source and a v0.4.0 GitHub Release are available. This is a library, not a hosted application or a drop-in replacement for the archived coding-agent application.

Agent OS

Multi-Agent Coding Orchestration System

A persistent orchestration layer for Claude Code and Codex. Coordinate agents, workspaces, approvals and scheduled execution from Feishu / Lark.

VIDEO · 84-second orchestration demoARCHITECTURE · Shown hereSOURCE EVIDENCE · Reviewed 283a624
  • Session lifecyclesrc/core/session-manager.ts

    Session keys, allowed transitions, native IDs and persistence rollback.

  • Agent dispatchsrc/app/collaboration-service.ts

    Collaboration tasks and routing between configured agents.

  • Scheduler recoverysrc/app/scheduler.ts

    Scheduled execution and interrupted-run recovery behavior.

Source-reviewed orchestration behavior. No production availability or customer deployment claim is made.

Agent OS in action

84 seconds of real execution: task dispatch, human approval, Claude Code implementation and a follow-up in the same native session.

Synthetic demo repository · Waiting time condensed · English synthetic narration · Private conversation sidebar cropped.

Transcript and verification

A task arrives in Feishu: build POST /sum and regression tests. The team lead dispatches a product agent, which creates a specification for human approval. Once approved, local Spec/Tickets guide the developer using Claude Code. The initial implementation passes 9 tests.

A follow-up limits input to 100 numbers: 100 are accepted and 101 are rejected. All 11 tests pass. The developer’s native session identifier was compared before and after the follow-up and remained unchanged. These are results from this demo run, not production metrics.

RAGOps Studio

Production-Oriented Agentic RAG Platform

A full-stack knowledge system with inspectable hybrid retrieval, governed document versions, tenant-aware access and source-grounded answers.

ARCHITECTURE · Shown hereSOURCE EVIDENCE · Reviewed b428307
  • Hybrid retrieval & ACLbackend/app/services/retrieval.py

    Permission-scoped snapshots, index validation and reranking.

  • Agentic workflowbackend/app/services/graph.py

    Routing, bounded searches and optional LangGraph integration.

  • Document versionsbackend/app/services/store.py

    SQLite store, immutable revisions and active-source governance.

Production-oriented reference implementation. Public source and synthetic examples are inspectable; production-model accuracy is not asserted.

Try the retrieval. Inspect the evidence.

Run client-side retrieval over a synthetic corpus. Example queries use BM25 + Dense + RRF; answers are labeled fixture replay.

Demo WorkspaceSafe demo
12 documents
Synthetic reference corpus · No customer data · No external AI calls

Examples: real BM25 + Dense + RRF, using precomputed embeddings. Free text: real BM25. Answers: fixture replay only.

Choose an example or run a query.

Every answer should have a trail.

Run a query to inspect its retrieved evidence.

hybrid-retrieval / fixture-generationRuns entirely in your browser