Engineer & independent researcher
Pengcheng Xu.
At Speak I work on infrastructure and agentic products for language learners. I’m also building AgentConnect, an open-source platform for teams and AI agents.
Selected workAbout
I work where systems engineering meets independent research. Since July 2026 I’ve been an engineer at SSpeak ↗, the AI language-learning app, working mainly on infrastructure and agentic consumer products for learners. I’m also building AAgentConnect ↗, an open-source platform where teams and multiple AI agents work together in shared conversations.
Before that I was at MMicrosoft from 2020 to 2026, most recently as acting Tech Lead on Fabric Eventstreams, where we moved the real-time streaming engine onto a unified compilation pipeline. Earlier, at EElastic, I built code intelligence on language servers. I studied computer science at UUSTC.
A finding · arXiv, 2026
Does a language server save tokens for coding agents? Usually not.
Everyone says semantic retrieval is cheaper than grep; almost nobody measures it at equal task success. I defined one metric, tokens-to-success, ran a five-arm ablation on Python and TypeScript repositories across three Claude models, and the answer came out conditional and usually negative. The practical implication is not “always use LSP” but an adaptive router keyed on task class, model capability, and lexical noise.
- +6% to +118%
- token cost added by the LSP on symbol-named localization
- 0–6%
- how often models chose semantic tools over grep when localizing
- 3 in 4
- multi-file renames a location-only LSP failed; grep solved them all
Questions I’m chasing
- How should a team and many agents share one conversation? AgentConnect
- Can an agent’s execution traces be compiled into zero-token scripts? AgentJIT
- When does semantic code intelligence actually pay for itself? The LSP study
- What is the right intermediate representation for a streaming pipeline? Eventstreams V3
- Can release notes be derived from repository deltas instead of written by hand? Changelogue
- How do you isolate untrusted, multi-tenant workloads without paying for it in latency?
Projects
-
agents
AgentConnect ↗
A platform for teams and multiple AI agents to collaborate in shared conversations and workflows. Connect agents to Slack, GitHub, and other tools, with shared context, permissions, and self-hosted execution.
-
jit
AgentJIT ↗
A background JIT compiler daemon for autonomous LLM agents. It analyzes agent execution traces and compiles stochastic, multi-step reasoning paths into deterministic, zero-token shell scripts.
-
tooling
Changelogue ↗
An agent-driven release-intelligence framework that tracks repository deltas and synthesizes human-readable release notes automatically.
Systems work
-
compilers
Eventstreams V3 Topology Re-architecture
As acting Tech Lead, re-architected the core real-time streaming engine onto a unified backend compilation pipeline and a high-performance IR/DAG framework. Serves Fabric customers globally; Microsoft documents a single eventstream at up to 200 MB/s.
-
codegen
Sentio.xyz Agent Orchestration
Designed and shipped a LangGraph supervisor-graph agent orchestration system; built an LLM processor code-generation agent with multi-chain support (EVM, Sui, Aptos, Move); owned the EVM tracer, debugger, and Solidity language server.
-
synthesis
NL → Streaming-Pipeline Agent
An LLM agent that compiles natural-language business constraints directly into deployable Fabric Eventstreams pipelines. Built at a Microsoft org hackathon.
Background
- Speak Engineering · infrastructure and agentic consumer products for learners
- Microsoft Technical Lead / Senior Software Engineer · Beijing → Shanghai
- Elastic Software Engineer, Codesearch (via the Insight.io acquisition) · Beijing
- YITU Technology Software Engineer Intern, Infrastructure · Shanghai
- USTC B.Sc. Computer Science & Technology, University of Science and Technology of China
Upstream open source microsoft/Trill tikv/tikv elastic/kibana