← Back to Home

Internship AI Toolchain

Two self-built AI automation systems — a market-weekly Agent and an internal AI-news curation platform

  • 2Self-built systems
  • 20+Incremental iterations
  • 0→1From zero to live
  • Go
  • Python
  • React
  • Agent Automation
  • RAG
  • Cron Jobs

During my internship I independently designed, built, and shipped two AI automation systems, closing the loop from requirement to deployment myself. Together they represent my AI-engineering delivery capability at the internship stage.

1. mkt-weekly · Market Weekly Report Agent

A scheduled fetch → clean → process → push market-information weekly-report system that turns the repetitive “manually browse, copy, and summarize market updates” job into an unattended agent pipeline.

  • Ingest: RSS + multi-source fetch, normalized raw info, idempotent dedup.
  • Process: classification, scoring/scoring (grader), intent understanding, cross-article period filtering (aggregating scattered daily info into weekly report dimensions).
  • Insight: an insight step extracts key changes; digest_engine assembles a readable weekly report.
  • Push: push_digest / weekly_push schedule delivery to IM, with a digest_watchdog stability guard and webhook_server.
  • Engineering: full tests/ (15+ cases across intent/classify/grader/digest/loop/watchdog), crontab/plist scheduling, and a complete docs lifecycle (design/decisions/runbooks/handoff) — stable for long unattended runs.

2. ai-cool · Internal AI-News Curation Platform

A web product + REST API that rebuilds the external aihot experience — “daily curated AI report + hot topics + categorized search” — on the internal network.

  • Three-tier cross-machine pipeline: collection (external RSS/WeChat MP/X) → normalize + idempotent → processing (internal LLM pipeline).
  • All core intelligence self-built: five-axis scoring, classification, curation judgment, event clustering (multisource reports of the same event cluster, only primary emitted to avoid spam), heat algorithm (multi-source count + exponential decay), and daily editor-style report generation.
  • 1:1 open API contract reuse: the 6 read-only endpoints + data model + cursor pagination + weak ETag/304 + rate limiting + 7-day window fully aligned with openapi.yaml.
  • Complete engineering: Go backend (modular internal/ + full unit tests) + React frontend (Feed/hot topics/daily report/detail page/knowledge graph) + deployment scripts (supervisor/nginx/systemd), with 32+ incremental design docs.

How it was built with AI

Both systems are a full practice of AI pair-programming + Loop engineering:

  • Every feature followed “Superpowers clarify requirements → solution → plan → staged implementation → test & acceptance,” captured as docs/superpowers/specs and plans.
  • Automation loops let the systems run themselves: ingest → process → push fully async, guarded by watchdog — truly unattended.
  • From 0 to 1 in production: mkt-weekly runs on a scheduled on-site task; ai-cool is deployed to an internal dev box providing API and frontend, walking the full AI-coding deployment path.