Engineering Work工程作品

Production platforms, self-service data products, and LLM systems — from Antares Capital, Summoner, and Epsilla. More on GitHub.生产级平台、自助数据产品，以及 LLM 系统 — 来自 Antares Capital、Summoner 与 Epsilla。更多项目见 GitHub。

• Enterprise Workflow Orchestration Platform

  TypeScript · Nest.js · MySQL · React
  Antares Capital, New York — Software Engineer

  Built an end-to-end workflow platform to replace multi-week, manually coordinated operations with a centralized, auditable system across MySQL, REST APIs, and a micro-frontend UI. A versioned template-instance model lets templates evolve without disrupting in-flight runs; group-based three-tier permissions enforce least privilege at process, step, and task scope. Schema-driven configuration auto-generates grids, forms, and filters, with alerts, audit history, health indicators, and email notifications backed by distributed scheduling.

• Governed Self-Service Data Product

  SQL · Cloud warehouse · RBAC
  Antares Capital, New York — Software Engineer

  A self-service analytics product that turns warehouse tables and views into business-friendly reports, so authorized users can discover, analyze, and save work without writing SQL. Governance and lineage rest on validated source mappings, UUID-based contracts, role-based access, lifecycle controls, and immutable report lineage — administrators can manage approved datasets without breaking existing reports or changing their meaning.

• Server-Side Pagination Re-architecture

  SQL · Redis · Microservices
  Antares Capital, New York — Software Engineer

  Re-architected a legacy data grid from client-side full loading to server-side pagination with database-driven SQL queries, pagination-optimized indexing, and Redis-based distributed caching. Reduced payload size from ~50MB to ~200KB and page load time from 10–30s to 1–2s, improving query latency under large datasets and enabling stateless services that scale horizontally across API servers.

• Multi-Agent LLM Evaluation Platform

  Python · GPT-4o · Asyncio
  Summoner, New York — AI Engineer

  An LLM-driven platform that ingests user context, generates domain-agnostic challenge questions with logical reasoning pathways, synthesizes AI personas, and orchestrates real-time timed challenge sessions. Answers are evaluated with GPT-4o structured logic extraction and embedding cosine-similarity scoring for automated partial credit. Includes a configurable multi-factor scoring system, cost-observability safeguards with real-time USD tracking and budget ceilings, semaphore-bounded async concurrency, and a state-machine protocol letting persona agents request peer hints through async message queues.

• Evaluation Analytics Pipeline

  SQL · PLS Regression
  Summoner, New York — AI Engineer

  A SQL-based pipeline collecting agent telemetry and experimental data at scale, powering statistical diagnostics and benchmarking. Includes a Partial Least Squares regression dashboard with cross-validated Q² metrics and VIP feature-importance scoring to identify performance drivers across agent configurations.

• Local RAG Research Agent

  LangChain · Ollama · RAG
  Epsilla (YC 23), New York — AI Product Strategist

  An Ollama-powered automated research system built on LangChain: multi-source search (DuckDuckGo, Tavily, Perplexity), local LLM summarization, YouTube transcript ingestion, configurable research loops, and SMTP delivery of concise, citation-backed reports — improving research efficiency by 27%.