Personal lab of Rohan Saraf — Senior Architect & Interaction Designer. This is where ideas about AI, agents, and interfaces become real software.
Rohan Saraf is a Senior Architect of Interaction Design at GE HealthCare, with 15+ years leading product design across complex systems. Outside of work, he runs icuboid — a personal AI lab where he builds tools that sit at the intersection of autonomous agents, product design, and self-hosted infrastructure.
The philosophy here is simple: understand why before building. Architecture first. Tight scope. Each project is a real experiment — deployed, running, and being used — not a demo or a proof-of-concept that lives on a slide.
The stack is deliberately heterogeneous: n8n for workflow automation, Supabase for persistence, Ollama + Forge for local model inference on Apple Silicon, and a self-hosted Hetzner server for everything else. The constraint is the feature.
A cowork-style local AI workspace on the Mac. Manages the full local model stack — Ollama + Stable Diffusion Forge — and provides an agentic chat environment with file tools, skills, MCP support, and a live HTML artifact preview panel. Built because no existing tool was designed with the agent loop as a first-class citizen.
"Lighthouse for AI agents." Submit any URL and get an Agentability Score (0–100) across 8 Agent Factors Engineering (AFE) principles — machine readability, status, chunking, handoffs, and more. A public research project asking: how well can AI agents actually operate today's software? AFE is original research coined here.
An ambient wellness Telegram companion that holds natural conversation and silently extracts Apple HealthKit-shaped State of Mind signals from what's said. A custom iOS SwiftUI bridge writes the extracted data directly into Apple Health. The core question: can unstructured chat become a reliable biometric signal?
A multi-character AI debate engine. Pick characters from history, fiction, or ideas — they argue, react, and build on each other's positions in real time. V2 added news-seeded debate topics via BBC RSS and Supabase-backed conversation memory with post-session summaries. V3 in backlog.
A Slack-native AI planning assistant. Mention @RIHANA in a channel and it pulls context from Supabase memory, reasons with GPT-4o, logs the interaction, and replies — all within Slack. Built as a persistent, contextual thought partner rather than a stateless chatbot.
A structured design-thinking companion with 11 chapters. Runs on Claude Sonnet and maintains persistent memory in Supabase — so it actually remembers what you've worked through before. Built for design practice, not generic Q&A. Phase 2 adds pgvector semantic retrieval.
Submit a name and get back a deep character profile — voice, worldview, quirks, blind spots — along with a generated avatar. Perplexity Sonar handles live research, GPT-4.1-mini synthesises the persona, and Pollinations generates the image. The whole pipeline runs in under 30 seconds.
Every morning at 7AM IST, six RSS feeds are pulled, each article scored by GPT-4.1-mini for relevance and quality, and only those scoring ≥6 make the cut. The digest arrives as a clean email via Resend. A small but genuinely useful automation that's been running for months.
An Agent-First OS prototype: an OS-metaphor interface for agent orchestration with a round-robin process scheduler, HITL (human-in-the-loop) interceptor for shell commands, and a cockpit dashboard for telemetry. Frozen as a portfolio exhibit — its supervision patterns are being absorbed into icuboid Studio as four named Agent-First UX principles.