Saved Self

Turn saved inspiration into small life experiments to discover who you want to become.

Comment

Inspiration

Many people save inspiring images, quotes, and ideas, but most of that inspiration never turns into real action.
Saved Self was inspired by that gap: a user may be moved by saved content, yet still not know what to do next.

What it does

Saved Self is a local AI workflow that turns saved inspiration into one small, realistic life experiment.
It first proposes several tentative possible selves, lets the user correct them, then designs a constrained, low-effort experiment, captures confirmation, and supports reflection.
The app updates the statement based on feedback while preserving the confirmed experiment and keeping all actions user-driven.

How we built it

We built a local web prototype with:

  • Python backend with HTTP routes and persistence
  • Strands Agents for the multi-stage workflow
  • Ollama local inference (saved-self-qwen3:1.7b)
  • SQLite state storage and deterministic validation

We used strict stage-level tool separation (read then write), state revision checks, and limit checks for time, budget, and location.
The frontend is a single-page interface showing each stage clearly, with explicit user confirmation before calendar export and visible simulated-feedback labeling.

Challenges we ran into

  • AWS Bedrock access was blocked by account-level allowlisting during setup, so we had to keep the local-model path as the reliable route.
  • Ollama tool schema rendering had compatibility issues and needed explicit JSON tool-template handling.
  • The model once issued write operations too early, so we added guards to block write actions unless the read result was actually available.
  • End-to-end quality still needs human semantic review: passing structural checks does not guarantee ideal phrasing or relevance in every run.

Accomplishments that we're proud of

  • A resilient stage pipeline with read/write boundaries and source-validation safety checks.
  • Constraint-aware experiment generation that respects user limits (minutes, budget, and at_home).
  • Reproducible acceptance flow with real local inference and persistence checks across refresh.
  • Full local test build, structured validation logs, and a public repository prepared for submission.

What we learned

A practical agent is less about prompt cleverness and more about contract safety: who is allowed to act, when, and with what evidence.
In practice, workflow reliability came from guardrails (state checks, schema validation, explicit user confirmation) combined with minimal, clear prompts.
Local-first architecture helped with privacy, speed of iteration, and predictable debugging.

What's next for Saved Self

  • Improve onboarding and inline guidance for first-time judges/users.
  • Expand behavioral edge-case coverage for failed/invalid model outputs.
  • Add usability improvements in the demo flow and clearer interpretation previews.
  • Add optional user account mode with stronger long-term consistency, while keeping local-first mode as the default privacy-first option.

Built With

Share this project:

Updates

Submission history