NightOwl

• 

  nightowl the bedtime companion

Inspiration

Bedtime for kids is often chaotic: overstimulation from screens, worries about tomorrow, and generic content that doesn't respond to how they actually feel.

We wanted to turn that moment into something calmer and more relational -- using AI as a gentle, always-available story partner rather than another feed to scroll.

The Gemini Live Agent Challenge's focus on redefining interaction was a perfect fit. We asked: what if an agent could listen like a caring friend, think like a creative director, and perform like a storyteller -- all in real time, across voice, text, and images? NightOwl is our answer to that question for kids at bedtime.

What it does

NightOwl helps kids wind down by turning bedtime into an interactive story they control.

A child like Alex -- a curious, imaginative 9-year-old -- talks to a friendly animated owl, shares how they're feeling, and describes the adventure they want: "A space story with my dog, but not too scary."

Using Gemini Live, NightOwl listens continuously, understands interruptions, and responds instantly with narrated scenes, generated illustrations, and gentle background audio. If Alex changes their mind -- "Make the alien silly, not mean!" -- NightOwl immediately adapts the story, visuals, and tone, reinforcing a sense of safety and agency while still feeling magical.

Parents get peace of mind that bedtime screen time is calmer and more creative, while kids get a consistent night-shift companion that remembers their favorites and builds on prior adventures over time.

How we built it

NightOwl is built as a live, multimodal agent on top of Gemini:

• Gemini Live API + GenAI SDK handle full-duplex audio so kids can speak naturally, interrupt, and correct NightOwl mid-story without rigid turn-taking.
• Interleaved output from Gemini lets us generate narration, inline images, and structured cues for background audio in a single response stream, so the visual story and spoken story stay in sync.
• A backend service on Google Cloud (Cloud Run) manages sessions, safety policies, and story state, including the child's preferences.
• Firestore stores story seeds, recurring characters, and parental settings so NightOwl can build multi-night arcs and remember what each kid likes.
• A web/mobile front-end renders an animated owl avatar, displays text and illustrations as scenes unfold, and streams low-latency audio back to the child.

We designed NightOwl's brain as a storyboard pipeline: Gemini doesn't just output text -- it emits a structured story spec (scenes, moods, visual beats) that the client uses to coordinate narration, imagery, and UI transitions.

Challenges we ran into

• Making interruptions feel natural required careful handling of partial transcripts and context so Gemini could smoothly resume the story from the exact emotional beat where the child cut in.
• Designing the interleaved output schema -- what belongs in narration, what becomes an image, what becomes a music cue -- took several iterations before the experience stopped feeling disjointed.
• Balancing creativity with safety for kids meant building comfort rails around topics, tone, and intensity, and giving parents clear controls without breaking the magic of a free-form story.

What we're proud of

We're proud that NightOwl feels like a character, not a command line. Kids can talk over it, change their minds, or get silly, and NightOwl gracefully adapts instead of failing.

We're also proud of how much happens in one continuous, multimodal loop: listening, planning, narrating, drawing, and adjusting -- all powered by a single Gemini-based agent rather than a brittle chain of disconnected calls.

What's next

• Run small pilots with families to test which story modes (calm, silly, educational) best help kids relax and sleep.
• Expand NightOwl's memories so it can reference past adventures and build ongoing worlds kids can revisit night after night.
• Explore sibling and parent co-play modes where NightOwl coordinates bedtime stories for more than one child at once.

Built With

• cloud
• fastapi
• gemini
• google
• python
• runfirestore