Kris — Your Personal AI Companion

One companion, one brain: a persistent voice-first AI that remembers context and coordinates Marketplace, navigation, work, and daily life through one trusted response authority.

Comment

Inspiration

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for Kris — Your Personal AI Companion

Inspiration

I did not start Kris because I knew how to build AI. I had no AI-development experience. I started because I was worried that forgetting an important detail at work could get me fired. A work-report helper gradually became something larger: one companion that could carry important details across work, messages, appointments, errands, navigation, and ordinary conversation.

Kris is not a Marketplace bot, work app, GPS helper, or row of feature buttons. Those are abilities. Kris is meant to be one persistent companion who knows which ability to use, keeps one identity, and carries decisions forward.

What it does

Kris is an Android-first, voice-first personal AI companion. The current project accepts voice, typed input, notifications, and device context. It uses GPT-5.6 through the OpenAI Responses API for natural conversation and structured extraction, with local fallback paths when cloud AI is unavailable.

The Build Week work tackles a difficult real-world problem: several Android listeners, services, and asynchronous callbacks could behave like separate versions of Kris. The one-brain architecture introduces one response authority, one pending-question authority, one speech gateway, generation-owned work, authoritative power epochs, and Living Continuity so a completed commitment stays completed across reminders, routes, and follow-up questions.

How I built it

I describe real behavior in ordinary language: Kris interrupted me, repeated an old subject, acted like two people, or turned herself back on. Codex investigates the live source and device evidence, identifies the competing owner or stale callback, proposes an invariant, implements the correction, writes deterministic tests, and gives me a build to verify in real life.

The implementation is a native Android/Java application using Android SDK 35, Android SpeechRecognizer, offline Vosk recognition, GPT-5.6 Terra and Luna through the Responses API, an encrypted Python phone-to-computer hub, owner-controlled local state, deterministic JUnit tests, and an auditable Git history.

Challenges

The hardest problem was not generating text. It was keeping one companion consistent across recognition engines, notifications, lifecycle recovery, cloud responses, local responses, location work, and text-to-speech.

The second challenge was honesty. Kris must distinguish what she heard, saved, inferred, can access, and does not know. A convincing answer is not a substitute for committed state.

The third challenge was continuity in development. Accepted, rejected, and unproven work must remain distinct so a later session does not restart from a simplified version and erase what was learned.

Accomplishments

  • Built one-brain ownership rules from real daily-use failures.
  • Added deterministic Off and Off-to-On race protection across listeners, background answers, seller notifications, and pickup radar.
  • Connected verified pickup closure to Living Continuity so tasks, questions, reminders, and route stops close together.
  • Passed 834 Android tests across 129 suites with zero failures, errors, or skips.
  • Passed all 25 encrypted hub tests.
  • Verified live GPT-5.6 Terra and Luna Responses API access without embedding the API key.
  • Built this as a non-technical creator by turning everyday feedback into engineering evidence.

What I learned

AI personality is partly an architecture problem. If several services can independently accept, mutate, and speak, the user experiences several personalities no matter how good the prompt is. Memory also has to be more than a summary: durable facts, active commitments, conversation context, history, and temporary observations need different lifetimes and provenance.

What's next

Kris will continue toward one companion that can help with daily work, messaging, navigation, appointments, opportunity analysis, and ordinary conversation without turning each capability into a separate app or identity.

Kris is an active personal project, not a finished commercial product. Release 6.09 remains the accepted phone build; the 6.21 demo work is shown honestly as a field-tested candidate rather than falsely promoted.

Built With

Share this project:

Updates