Home view
Share Card to share progress online

Devpost Submission — Delulu

About the Project

Inspiration

Gabby Beckford (@packslight) has spent years telling her community of 1M+ ambitious women that "Delulu is the solulu" — that delusional confidence is the key to going from dreaming to doing. But there's always been a massive gap between seeing an inspiring Reel and actually taking action.

Gabby said it perfectly:

"There's still this massive gap between inspiration and action. They stay stuck — waiting for permission, waiting to feel confident."

That hit hard. I realized the problem isn't motivation — her community has plenty of that. The problem is translation. Turning a bold, terrifying dream ("negotiate a six-figure salary", "travel solo to Japan", "start my own business") into something you can actually do today, in 15 minutes, without feeling overwhelmed.

That's what Delulu does. It takes the impossible and makes it inevitable — one micro-step at a time.

What It Does

Delulu is an AI-powered confidence coaching app built natively for iOS. Here's the loop:

Dream big — Enter any goal, no matter how "delusional"
AI breaks it down — Apple's on-device Foundation Model (a 3B parameter LLM running entirely on your iPhone) generates a phased, personalized action plan — grounded in your real context (location, time, interests, occupation)
One task per day — The AI picks the single best task for right now based on time of day, deadlines, goal urgency, and category variety. You never see the overwhelming full plan
Micro-steps — Each daily task gets broken into 3-5 bite-sized actions (2-10 minutes each). Check them off, swipe to complete, earn XP
Level up — Your mascot companion evolves through 4 stages (Dreamer → Starter → Achiever → Legend) as you prove the doubters wrong
Celebrate — Confetti, streaks, share cards. Because every win deserves to be loud

How I Built It

100% native iOS — SwiftUI, SwiftData, WidgetKit, and Apple Foundation Models. Zero external APIs. Zero cloud dependencies. Everything runs on-device for complete privacy.

The hardest engineering challenge was making a 3B parameter on-device model produce consistently structured, high-quality output. I developed a three-layer context architecture:

Layer 1 — Profile Memory: An AI-synthesized summary of the user's identity, regenerated only when they update their profile. Natural language, not raw fields — because the model reasons better over prose than structured data
Layer 2 — Goal Digest: A daily AI-generated summary of the user's goal landscape (momentum, deadlines, neglected goals). ~80 tokens of distilled understanding, plus a live delta of what changed since the digest. This replaced a 400-token verbose data dump
Layer 3 — Master Context: Assembled on-demand for each AI session — time, location, profile memory, and goal context. Every interaction is grounded in reality

I also built custom Tool Calling with MapKit (real places near the user) and a Date Calculator (deadline feasibility), so the AI can suggest "Visit Gold's Gym on 5th Avenue" instead of "go to a gym."

The DailyTaskEngine uses phase-aware ordering and AI-powered selection to intelligently rotate between multiple goals. It considers time of day (focused tasks in the morning, lighter ones at night), category variety, deadline urgency, and goal neglect — then explains its reasoning in the mascot's confident voice.

Challenges

On-device model limitations — The 3B model has a small context window and can hallucinate. I built aggressive grounding rules, tool-based fact injection, and the three-layer context system to keep outputs accurate without making them verbose
Structured generation reliability — @Generable types sometimes fail deserialization on complex nested objects. I implemented a two-pass strategy: try with tools first, retry without tools on failure
Gamification without distraction — Gabby wanted gamification as a seasoning, not the dish. The XP/mascot system enhances completion feel without adding cognitive load to the core "one task per day" loop
Making AI feel personal, not robotic — The AI speaks as "Lulu, your confidence coach" — warm, bold, unapologetically optimistic. I tuned the identity prompt extensively so it never says "consider" or "you might want to." It says "Lock this in." "Light work for you." "This is your moment."

What I Learned

On-device AI is ready for production — but you have to architect around its constraints, not fight them
The best gamification disappears into the product. Users shouldn't feel "gamified" — they should feel celebrated
Gabby's audience doesn't need another to-do list. They need a hype woman who also happens to be organized