Zhou Zhihao
Building evidence-grounded AI systems, reliable backend services, and practical developer tools.
Records every agent action as evidence, checks each claim with a deterministic engine, gates consequential actions behind human approval, and seals the run into a versioned, SHA-256-hashed AI Work Receipt.
237 tests · CI green · 7 MCP tools · 3 MCP resources
Ingests documents, extracts claims, and verifies each against retrieved evidence — deterministic checks are the guardrail, with optional semantic verification that can never override a contradiction.
117 tests • 1 intentional xfail
A source-aware tender workspace that extracts page-referenced requirements from PDFs, answers grounded questions with cited sources, and assembles a submission package with an evidence vault and bid tracker.
177 tests · 0 failures · 71s demo
A local RAG prototype that chunks a PDF, retrieves the most relevant passages with semantic and keyword scoring, and answers questions through an Ollama model grounded in the retrieved source pages.
Local RAG · multilingual embeddings · source pages
Accepts telecom SMS and data-usage events over HTTP, validates them at the boundary, buffers them in memory for async batched flushes, and writes them to PostgreSQL in idempotent bulk inserts chunked to respect the parameter limit.
52,247 events/sec · 100k events · 0 failed
A Java 21 library that sanitizes untrusted HTML by parsing it into a DOM with jsoup, walking the tree to apply an allowlist policy, stripping event-handler attributes, validating URLs, and emitting clean HTML alongside a SecurityReport.
363 tests · 0 failures · 0 errors · BUILD SUCCESS
I build practical software projects that solve concrete problems, with a focus on AI applications, backend systems, data processing, and automation.