商业分析智能体Business analysis agent
工程案例Engineering case study
围绕 147 万条公开交易明细，将业务语义、多智能体调查、确定性分析与经营报告串成完整工作流。Connect business semantics, multi-agent investigation, deterministic analysis, and reports over 1.47 million public transaction rows.
↗
+10 pp驱动召回率 · 同输出、10 案例消融Driver recall · same outputs, 10 cases
Driver Recall@K 从 90% 到 100%；只比较同一批分析结果的选择策略，不能解释为营收提升。Driver Recall@K rises from 90% to 100% when selecting from the same analysis outputs. This is a selection-policy comparison, not a revenue increase.
查看项目案例Read the case study →GitHub ↗
内容生产与长视频再生产Media production and video repurposing
内容生产工程原型Media production prototype
以转写和时间戳为依据筛选长视频片段，连接质量门、人工审核、FFmpeg 导出与阶段恢复；同时保留生成式内容生产链路。Ground long-video clips in transcripts and timestamps, then connect quality gates, human review, FFmpeg exports, and recovery alongside a generative media pipeline.
实现与交付Implementation & delivery
FastAPI 任务接口、PostgreSQL Checkpoint、FFmpeg 导出/批量 ZIP、供应商适配器、MinIO 恢复与确定性测试。FastAPI task endpoints, PostgreSQL checkpoints, FFmpeg exports and batch ZIPs, provider adapters, MinIO recovery, and deterministic tests.
53 tests多模态范围 CI · 包含真实 FFmpeg 媒体检查Multimodal-slice CI · includes real FFmpeg media checks
媒体测试使用本地生成样例，ASR 与规划使用固定替身；不是实际模型质量或业务收益的验证。Media tests use generated fixtures with recorded ASR/planner doubles; they do not establish model quality or business impact.
查看项目案例Read the case study →GitHub ↗
领域 SFT 与 LLMOpsDomain SFT and LLMOps
领域微调工程Domain fine-tuning engineering
连接源级数据隔离、assistant-only 两阶段 LoRA 训练、质量评测和 vLLM 双版本部署。Connect source-level data isolation, assistant-only two-stage LoRA training, quality evaluation, and dual-version vLLM serving.
实现与交付Implementation & delivery
数据治理与泄漏审计脚本、两阶段 LoRA 训练入口、合并与 vLLM 双版本服务、评审器和汇总快照。Data-governance and leakage-audit scripts, a two-stage LoRA trainer, model merging and dual-version vLLM serving, a quality judge, and an aggregate snapshot.
Agent 强化学习Agent reinforcement learning
RL 工程原型RL engineering prototype
实现无答案泄漏的 SQL 环境、Agent Lightning／veRL GRPO 接口，以及权重变化与固定预算对照验收。Implement a Gold-isolated SQL environment, Agent Lightning/veRL GRPO integration, and checks for weight changes and fixed-budget controls.
实现与交付Implementation & delivery
只读 SQLite 环境、真实框架配置接口、Checkpoint 导出、张量变化检查、配对统计及证据门禁。A read-only SQLite environment, upstream-framework configuration integration, checkpoint export, tensor-delta checks, paired statistics, and an evidence gate.
多模态模型评估Multimodal evaluation
多模态研究Multimodal research
固定问题和候选回答，只改变视觉证据；用 8 个模型、4 类因素检验静态分数相近时的干预行为差异。Hold questions and responses fixed, edit visual evidence, and compare intervention behavior across eight judges and four factors.
↗
15.1 ppAttribute 静态平分下的 RA 差异RA gap at an exact Attribute static tie
Attribute：两模型静态准确率均为 80.5%；RA 分别按各自原图正确子集计算。此汇总对比与上面的 Molmo 单例使用不同模型。Attribute: both judges score 80.5% static accuracy; RA uses each judge’s own base-correct subset. This aggregate comparison uses different models from the Molmo example above.
查看项目案例Read the case study →GitHub ↗
