AI orchestration developer and front-end systems engineer.
I design the connective tissue between models, tools, agents, codebases, and users.
My work sits at the point where software systems need to stay understandable while becoming more intelligent: context engineering, knowledge graphs, agent workflows, developer tooling, and front-end interfaces that make complex behavior feel clear.
AI context infrastructure
Memory systems, repository-aware knowledge graphs, agent workflows, tool routing, and evaluation loops for AI-assisted development.
Front-end systems engineering
Interfaces and component systems that survive real product constraints: accessibility, localization, timezones, testing, state flow, and edge-case behavior.
Developer experience
Build tooling, CI parity, reusable workflows, documentation, and automation that makes teams faster without hiding important complexity.
Repo-aware knowledge graphs for AI assistants
I adapted Graphify for a large multi-package monorepo where the upstream assumptions did not fit the host environment. The useful work was not a direct install; it was the reconciliation layer around it:
- package-level graph storage instead of a single repo-root graph
- sequential, resumable extraction for an assistant without background workers
- graph routing by closest package path
- tracked Git hook automation for affected-package rebuilds
- deep, directed graph defaults
- setup and doctor scripts for local verification
- Memory Bank alignment so high-level summaries stay in the root memory layer while graph execution state stays package-local
Persistent context for AI-assisted development
I worked on Memory Bank and Cline workflow rules that help AI assistants preserve project context across sessions, follow repository conventions, and avoid repeatedly rediscovering the same architecture.
Production front-end systems
I have built and hardened component systems with keyboard accessibility, timezone correctness, localization, Cypress coverage, Storybook documentation, and careful state modeling.
The bottleneck for useful AI tooling is often not model capability. It is context quality.
A capable model still performs worse when it has to spend its working memory reconstructing architecture from scratch. I like building the systems around the model: the memory, graph structure, routing rules, verification steps, and human-facing interfaces that turn intelligence into something durable.
The theme I keep returning to:
Flow under control.
Make hidden state explicit. Route context correctly. Verify the boundaries where systems drift.
- GitHub: @grazomarin
- Website: coming soon