// dyeyc.dev
AI engineer at SOFI AI Tech Solutions and 4th-year CS student at PCU Bulacan, class of 2027.
I build AI products end to end: agents, RAG and the evals that show they work. My rule is that a 200 is not proof. Every project below ends in a measured result, not "it runs".
const status = {
location: "Philippines π΅π",
building: "relaydesk β AI support for small businesses",
certs: ["Oracle GenAI", "AWS"],
openTo: "AI engineer / forward deployed engineer roles, 2027",
};
ποΈ relaydeskAI customer support as a widget plus a staff inbox. It answers from the business's help center with citations, and when the help center has no answer it says so and opens a ticket. A co-pilot drafts the reply; a person sends it. One help center per business, self-serve setup, and an MCP server so agents can work the inbox. retrieval hit@3 55/60 Β· ticket precision 1.00 Β· Langfuse experiment gate on every PR |
βοΈ auto-learnFix your sentence, and learn the word that fixed it. An ESL writing tool where the better word is withheld server-side until you open its card: the replacement text does not exist on the client until the teaching does. Now with a speaking partner. pnpm monorepo Β· shared Zod contract Β· evals Β· Playwright e2e |
π rag-eval-harnessYour RAG got worse. Your tests stayed green. A regression gate for RAG pipelines: score a golden set, keep every run, and fail CI when retrieval or answers slip. A weaker retriever cut context recall 0.880 β 0.740 and the gate exited 1. on PyPI Β· keyless CI mode Β· baseline diff view |
β closerA post-ship agent. Paste a merged PR: it marks the Linear issue Done, emails whoever asked, then re-reads both apps. A 200 from Gmail is not enough; the message has to show up in sent mail. Ambiguous matches mean zero writes until a human picks. hackathon build Β· verifier re-reads every write Β· claims checked in CI |
I fix bugs in the AI tools I build on. Merged pull requests, counted live:
| Project | What it is | Merged |
|---|---|---|
| kyegomez/swarms | Multi-agent orchestration framework | |
| Asymptote-Labs/agent-beacon | Memory layer for AI agents | |
| mcp-use/mcp-use | Full-stack framework for MCP servers, clients and agents | |
| hydra-db/selfgraph | Versioned knowledge graphs for AI agents | |
| traceroot-ai | Open-source observability and debugging for AI agents |
AI
Languages & Frameworks
Data & Infra



