A LangGraph agent answering SEC-filing questions via PageIndex agentic retrieval at $0.0127 per question, served over MCP on Google Cloud Run. A Progent guardrail cut prompt-injection attack success from 63% to 0% over 108 AgentDojo-template attacks and blocked no legitimate email.一个 LangGraph Agent,用 PageIndex agentic 检索回答 SEC 财报问题,每题成本 $0.0127,通过 MCP 提供服务并部署在 Google Cloud Run。Progent 护栏在 108 次 AgentDojo 模板攻击上把 prompt injection 攻击成功率从 63% 降到 0%,且没有拦下任何一封正常邮件。
Live demo ↗在线演示 ↗Projects项目
CanPlan — RAG Task Planner
AIAIBackend后端A RAG backend on AWS that turns a daily-living goal into grounded, step-by-step plans, each citing its source. Three measured eval rounds on CanPlan's step-generation service ended at 100% valid citations and a reading level cut from grade 7.2 to 6.0 — and showed the safety guardrail still over-flags (96% → 89% of plans).AWS 上的 RAG 后端,把一个日常生活目标变成有依据、分步骤、每一步都注明出处的计划。对 CanPlan 的步骤生成服务做了三轮实测评估,最终引用有效率 100%,阅读难度从 7.2 年级降到 6.0——同时也表明安全护栏仍然会过度标记(96% → 89% 的计划)。
VAD Model Compression
Machine Learning机器学习Benchmarked PTQ, QAT and distillation on a CRDNN speech-detection model. The sweet spot is 0.185 MB (−57%) at 1.4 F1 points below the FP32 baseline; the smallest model, 0.053 MB, costs 2.6 points. Dynamic quantization was a no-op: only 1.2% of the weights are nn.Linear.在一个 CRDNN 语音检测模型上对比了 PTQ、QAT 和知识蒸馏。最佳平衡点是 0.185 MB(−57%),F1 比 FP32 基线低 1.4 个点;最小的模型 0.053 MB,代价是 2.6 个点。动态量化几乎不起作用:只有 1.2% 的权重是 nn.Linear。
Distributed Order Processing System
Backend后端An e-commerce order backend split into small services that never oversell stock, even under load — 599 concurrent orders on 200 units settle to exactly 200. It sustains ~1,884 req/s at 56 ms p95, and a connection-pool tuning round cut p95 21% (140 → 110 ms).一个拆成多个小服务的电商订单后端,在高负载下也不会超卖——200 件库存上的 599 个并发订单最终恰好成交 200 单。它能持续承载 ~1,884 req/s,p95 为 56 ms;一轮连接池调优把 p95 降低了 21%(140 → 110 ms)。
Distributed Book Recommendation
Backend后端A book recommender that chews through 30 GB+ of public library data into a 50,000-book catalog, then fans queries out across 26 workers / 27 partitions — figuring out which steps actually benefit from going parallel.一个图书推荐系统,把 30 GB+ 的公共图书馆数据处理成 50,000 本书的目录,再把查询分发到 26 个 worker / 27 个分区上——并弄清楚哪些步骤真正能从并行中获益。
Movify — Movie Social Network
Backend后端A full-stack movie social network where people register, follow each other, and post reviews — then like and vote on both movies and reviews. A React + Redux front end talks to an Express / MongoDB API that handles sessions and the social graph, and proxies live movie data from the TMDB API.一个全栈电影社交网络:用户可以注册、互相关注、发影评,还能给电影和影评点赞、投票。React + Redux 前端对接 Express / MongoDB API,后者负责会话和社交关系,并代理 TMDB API 的实时电影数据。