Software developer · builder · creative technologist

I build web applications, backend systems, and AI tools where the human stays the author and the machine stays accountable.

I'm a Computer Science student at Calvin University, working at the intersection of software engineering, AI systems, and creative technology.

• Reflective mind-mapping (research, AIToolsLab/writing-tools) — An AI writing tool that helps you externalize your own thinking into a map without ever authoring structure for you. The AI questions and mirrors your words back; every card, hierarchy, and connection is user-authored, with the guarantees enforced in code and hardened by a property-based fuzz harness that attacks the system with an adversarial mock LLM. Co-first author on the resulting paper, under review at a major HCI venue. On the product side, I migrated the platform's authentication from Auth0 to Better Auth — sessions, Google OAuth, protected endpoints, and the OAuth device authorization flow (built because a Word add-in taskpane can't rely on browser cookies).

• Duet / Local Dual Agent — A local multi-agent coding orchestrator. Claude Code and OpenAI Codex implement and review each other's work in isolated git worktrees, behind human plan and merge approvals. Persistent Node/TypeScript service, versioned HTTP API, SQLite persistence, and a web dashboard with a native tool-calling manager agent (Groq / Gemini / OpenAI-compatible). 200+ tests.

• Nhyira OS — A personal creative operating system bringing together software, writing, media, and interaction design. Vite + React SPA with scroll-driven motion, Hono API, deployed on Cloudflare Workers.

• Trust model judgment; verify model authority. In Duet, agents propose — approving, merging, and running are human actions, enforced by the backend, not by prompts.
• AI can reveal its thinking; it can't hand you an artifact to rubber-stamp. The mind-map system shows the writer exactly what the AI is tracking and waiting for — but nothing reaches the map unless the user says it in their own words and confirms it.
• Enforcement in code, calibration in config. Guarantees live in tested code paths — including property-based fuzzing against an adversarial LLM — not in system prompts. If a constraint matters, there's a validator behind it.

• Local Dual Agent
  An orchestration workspace for collaborative AI coding workflows.

• First Thought
  A browser-based word-association game exploring language, pace, and reflection through play.

• Clicker Challenge
  A 2D reaction-based shooter game built with Python, Tkinter, and Pygame.

I work as a writing consultant at my university's Rhetoric Centre — the same non-directive philosophy that shapes the AI tools I build. I edit video, led a multimedia team for three years, and write about what I learn.

I've also earned international STEM recognition, including a Gold Medal at the VANDA Science Global Finals and medals across Singapore math olympiads.

Explore the fuller story at nhyiramante.com.