My Role
- AI Agent Intern at ByteDance on the strategy side of Dreamina (即梦) and CapCut (剪映), June – September 2026.
- Designed, shipped, and operated the runtimes and data pipelines behind the team’s competitive intelligence, and the weekly briefs built on top of them.
- Reports written by these agents, including the weekly AI Film & IP brief, were delivered directly to Zhang Nan (Kelly Zhang), former CEO of Douyin Group and ByteDance China and now head of the CapCut business. screenshot ↗
Workflow Runtime (Strat Market Radar)
A scheduled runtime that coordinates concurrent API, search, and browser tasks and writes every new competitor event to a shared monthly Feishu sheet, with tabs for X, WeChat official accounts, TechCrunch, official sites and third-party channels, and AI film & IP news.
- Cold starts. Every daily run starts from a versioned release package with a SHA-256 manifest, so it never depends on chat history or leftover files. It pulls X, WeChat, and TechCrunch through primary APIs with per-account fallbacks, over a 24-hour main window plus a 72-hour late-recovery sweep.
- Safe retries. WAL-backed SQLite checkpoints and idempotent writes mean a failed or repeated run picks up where it stopped without writing anything twice.
- Correctness gates. If source coverage or pagination completeness drops below 95%, the production write is blocked. Unknowns stay visible as warnings instead of being reported as “no update”, and items that fail to parse go to a manual-review tab.
- Verified writes. Each run builds a deterministic write plan, writes, reads the columns back, and uploads an audit bundle. Zero duplicates per tab and zero post-write mismatches are completion criteria.
- Weekly crawl layer. A Firecrawl-based runtime covers 49 official-site entry points across 19 companies plus ~70 AI film & IP sources, gap-filling through RSS, sitemaps, and JSON-LD, with fingerprint-based dedup that keeps multiple changelog events per page.
- Results. On 42 labeled events, recall went from 32% to 95% with zero duplicates.
Feedback Platform
- Owned and shipped Dreamina’s Python feedback platform: 100K+ posts and comments on frontier video models, including Seedance 2.5, at 1K+ API calls per week.
- Bounded, source-aware concurrency, adaptive retries with backoff, and budget controls keep multi-source ingestion reliable and inside its API budget.
- Classifies and ranks feedback by relevance, sentiment, theme, and engagement, cutting synthesis time by about 90%.
Data Pipelines & Weekly Briefs
- Incremental Python/SQL pipelines across 30+ AI products, joining Sensor Tower revenue and DAU with web traffic, launch, and funding data through normalized schemas, per-source watermarks, and replayable checkpoints; data preparation time dropped about 75%.
- Weekly AI creative-tools industry brief: usage trends across ~35 products, plus track, funding, people, and product news.
- Tuesday competitor brief for the Dreamina content meeting, with rule-based growth attribution: any metric up more than 5% week over week gets a primary and secondary cause, a confidence level, and sources.
- Bi-weekly Web and App MAU/DAU benchmark across ~50 products for CapCut, built on a source registry of ~150 competitor products and ~110 media, investor, and creator accounts.