Official Didi technology account
DiDi's HUATUO project presents a kernel observability framework for agent sandboxes using eBPF and dynamic tracing to capture system-wide metrics, exceptions, auto-tracing, continuous profiling, and hardware faults, correlating kernel events with container identities for root-cause analysis in cloud-native and AI infrastructure.
At Apache COC Asia 2026, Didi’s engineers presented three deep‑dive sessions covering the evolution of massive‑scale object storage with Ozone, cost‑effective lakehouse migration using Iceberg and real‑time ingestion, and AI‑friendly enhancements to the open‑source Seata project, detailing architectural choices, performance gains, and open‑source contributions.
The article interviews three newly appointed core contributors—Ray community member Fan Shilun, Apache Auron committer Gui Hua‑wen, and Linux kernel maintainer reviewer Li Nan—who describe how they turned business pain points into open‑source contributions, the skills they gained, and actionable steps for engineers to start and grow in open‑source communities.
The paper introduces FAST, a parallel reinforcement‑learning sampling framework for autonomous‑driving that decouples individual episode termination from global resets via Dynamic Parallel Sampling Alignment and Scaled Mask‑Padding Optimization, achieving up to 9.08× higher sampling throughput and up to 2× faster training while preserving zero policy loss.
Didi's International Business Group built an LLM‑driven quality‑inspection platform for Spanish and Portuguese support across ride‑hailing, food delivery, and finance, using three pipelines—intent verification, compliance assessment, and VOC trend analysis—that boosted intent accuracy to 86%, compliance accuracy above 90%, and cut manual reporting time from hours to minutes.
This article analyzes the latency bottlenecks of large language models in multi‑round AI Agent scenarios, introduces SpecForge‑based speculative decoding and Unified Sequence Parallelism (USP) techniques applied to the EAGLE-3 model, and presents benchmark results showing over two‑fold Accept‑Len gains and 35‑44% reductions in P95 token‑level latency while enabling 128K context training on an 8‑GPU node.
The article details the creation of the didi-ride-skill for OpenClaw, explaining how a single voice command triggers a full ride‑hailing workflow, the underlying MCP toolset, engineering trade‑offs such as file splitting, attention handling, cron isolation, key management, testing strategies, and future roadmap.
The STAPO (Spurious‑Token‑Aware Policy Optimization) algorithm, introduced by Tsinghua University's iDLab and Didi's Deep Sea Lab, tackles policy‑entropy instability and performance oscillation in reinforcement‑learning fine‑tuning of large models by mathematically analyzing token collision probability, defining spurious tokens, and applying a Silencing Spurious Tokens mechanism that yields state‑of‑the‑art results on multiple math‑reasoning benchmarks.
Huatuo, the open‑source deep‑observability platform backed by Didi, now supports real‑time monitoring of MetaX GPUs, offering detailed hardware metrics via Docker or Kubernetes deployments and exposing them through a /metrics endpoint for cloud‑native AI and operations use cases.
On November 3, Didi Open‑Source presented its ecosystem and four flagship projects—XIAOJUSURVEY, HUATUO, MPX, and KnowStreaming—to over a hundred Zhejiang University software students, sharing insights on enterprise‑grade open‑source practices, cloud‑native observability, cross‑platform development, and the role of open source in talent cultivation.
