portfolio-orqys

Portfolio narrative for Orqys, "AI-native engineering ops layer that converts tickets into production-ready pull requests via multi-agent orchestration." Status: Building.

Key extractions

• Thesis: "the mechanical parts of the job" (reading context, boilerplate, tests, PR descriptions) are compressible. Goal is a deployable PR, not a code suggestion.
• Agents: Context, Planning, Code, Review, PR Synthesis. Each scoped, shared context window, DAG orchestration for parallelism where dependencies allow.
• Key decision: multi-agent over monolithic — a single LLM call can't reliably handle ticket→PR end-to-end.
• Key decision: MCP for tool integration — no custom adapters for linters/type checkers/test runners.
• Hard problem: context window management — real codebases have thousands of files; retrieval uses semantic search + structural analysis (imports, call graphs) to build focused windows.
• Hard problem: ensuring generated code compiles — review agent runs sandboxed validation before PR.
• Philosophy quote (load-bearing): "The goal isn't to replace engineers — it's to compress the mechanical parts of the job so more time goes to design, architecture, and the genuinely hard decisions."