Forecasting + Agent delivery
Hot-period forecasting & Agent Claw
Algorithm intern / Method and agent workflow · May–Aug 2026
I matched short viewing histories to comparable titles, used / to estimate end-state metrics, and connected the forecasts to a Feishu Bot for daily departmental use.
- My work
- implemented the forecasting method and connected it to a live Feishu Bot with deterministic constrained intents, preflight, tool calls, JSONL recovery of successful sessions, and confirmed human feedback.
- Outcome
- on days 1–3 after launch, standalone Top5/q35 forecasts versus the previous baseline plus manual correction improved mean for traffic (plays and main-content consumption) and membership revenue by 9 pp, and advertising separately by 11 pp. The model formally replaced the previous workflow and supplies daily forecasts through the team and Claw.