Mengyun Wang

Five agents — Planner, Safety, Coder, Browser, Verifier — run by one orchestrator that owns every state change. Plans are scored twice, by the Safety agent and by fixed rules, and anything over the line waits for a person before it runs. Steps retry, the Verifier can send work back once, and every event is in the database before it reaches the UI, so a task survives a restart and can be replayed. Code runs in a locked-down container; web fetches are checked against private networks. Runs end-to-end with no model keys.

An OpenAI-compatible gateway and operations layer for model engines. Requests move through authentication, rate limiting, a prefix-response cache, warm-pool coordination, adaptive backend routing, and dynamic micro-batching before reaching an upstream engine. The console exposes chat, routing health, cache and batch state, model governance, and canary, rolling, and blue-green release controls.

The engine layer beneath the serving platform: Ray Serve ingress drives vLLM’s AsyncLLMEngine with continuous batching, PagedAttention, prefix KV caching, bounded request concurrency, and OpenAI-compatible SSE. A Hugging Face baseline and async load generator measure TTFT, tail latency, tokens per second, and goodput; Prometheus reads vLLM and NVIDIA DCGM metrics for GPU and cache visibility.

An engagement prediction system taken from notebook to something operable: leakage-safe features, cost-aware model selection under a paired-bootstrap tie test, an operating point chosen against a traffic budget, then introspection and drift monitoring around it.

A growing question bank for machine-learning-engineer interviews — ML coding, theory, LLMs and agents, ML systems, MLOps, recommender systems, AI safety, multimodal and behavioural — filterable by category and difficulty. Reference answers give an answer framework, key points, common follow-ups and further reading; anything unclear can be handed to Claude or ChatGPT for a walkthrough.

Kuaishou injected uniformly random videos into live feeds for two weeks. Testing on both exposure mechanisms across the identical window isolates exposure bias from drift — ranking survives it almost intact, calibration does not.