Engineer · Researcher · Builder of agent systems
I build systems on top of long-running agents — marketing automation, autonomous research, agent control planes. I care most about the unglamorous parts that quietly hold everything together: observability, token economics, skill evolution, and shared memory.
Operators only ever speak to the controller. Everything else — email blasts, ad creative, video, TTS — is handled by sub-agents that own their own domain. The whole platform is built so a single sentence can become a campaign.
The site you're reading. Hugo with a hand-rolled theme, ~400 bilingual long-form pieces, ~1500 matplotlib figures pushed through a homegrown pipeline to Aliyun OSS. Designed to feel like a magazine, not a blog template.
A control plane for long-running research agents that read papers, run experiments, and write reports. The design hinges on a shared memory architecture plus harness-based skill evolution — failed attempts compress into lessons that feed the next run.
An autonomous pipeline that takes a question and returns a finished report. The writer sub-pipeline is decomposed into small, reentrant agents tied to a skill-evolution loop — every failure feeds back into the harness and shared memory, raising the floor each run.
A live dashboard for agent operations: session timelines, rate-limit budgets, and token economics across providers. The premise is to treat agents as real systems with cost and failure modes — not as a prompt.
File-based persistent memory across Claude Code sessions. User profile, feedback, project state, and references each get their own type. MEMORY.md acts as an index; concrete fragments load lazily on demand.
From PAI-Designer visual modelling to EAS elastic inference — covering every critical node of a machine-learning production chain.
From model selection to agent orchestration — translating Bailian's raw capabilities into shippable product surfaces.
Provisioning agent systems with Terraform: a plan-validate-apply loop that turns natural-language intent into auditable, reversible real-world resources.
No projects match this filter.
How should a long-running agent split its budget across providers? Which steps are worth the most expensive model?
Which failures can compress into reusable skills, and which can only be discarded?
When multiple agents share one memory, how strict should the type system be?
The last mile from agent demo to shipped product — observability, replay, governance, user trust.