Coo

• 

  Coo is live with Meal prep automation (under beta-testing with 2 users)

Coo — a voice-first AI assistant that takes the invisible mental load of caretaking off a parent's plate: starting with meals, groceries, and (soon) scheduling. I also tried a fun experimental feature to engage kids with parents without the need of screen time.

Inspiration

Parenting runs on an invisible mental load — deciding what's for dinner again, keeping the pantry stocked, planning the week, remembering everyone's schedule. Most apps are reactive: they hand you one more thing to manage. We wanted the opposite — an assistant that quietly does the work. And because a caretaker's hands are usually full, it had to be voice-first.

Our core bet: the two heaviest, most repeatable loads are meal prep and scheduling — automate those well and you give real time back. That's Coo — Caretaker's Operations Officer. Not just for moms: anyone holding down the operational load of caring for a family deserves the same relief.

Cherry on top — I'm a Chithi myself (the Tamil word for a mom's younger sister, usually the family member who steps in as a second caretaker) to a toddler, and that caretaking journey is what inspired this product in the first place.

What it does

Coo is a voice-first assistant that works as a system of agents taking care of everyday workflows that save a caretaker's time and mental load. Today the meal agent, Tuesday, is the star:

• Learns your family through a conversational onboarding (dietary needs, who you cook for, how you shop).
• Generates a 7-day meal plan that respects dietary restrictions, lets you swap any meal for a fresh AI idea, or edit it manually — type your ow and ingredients directly, no AI required.
• Reads your fridge/pantry from a photo and keeps a live inventory.
• Builds a shopping list, splits it acroand remembers your brand and storepreferences.
• Adds items to your Instacart cart throng in your own logged-in browser.
• Proactively checks in after each meal's usual time — did you actually eat what was planned, or something else? — fully opt-in per meal, with youre windows, not a one-size-fits-allschedule.
• Keeps a persistent, scrollable chat hi with Tuesday is saved and browsable, notlost the moment you close the app.
• It's agentic — a single request can tran my week and add what's missing to myorder").

Voice itself is evolving past simple turn-taking: alongside the shipped tap-to-talk flow, we're piloting a realtime conversation mode where speech- run continuously via a streaming model,while reasoning stays entirely in Tuesday's own tool-calling loop — so you can talk over Tuesday mid-response and it actually stops and lg from mic volume.

The scheduling agent, Sunday, is next inerimental agent DJ which helps parentsengage young kids and toddlers, by jamming on the favorite beats and mixing it up with their own themes, vocals and characters — a Sing & Play mo with your kid over familiar tunes, withmore ideas to explore based on user feedback (as engaging kids is another huge opportunity).
How we built it

• iOS app: SwiftUI, project managed with XcodeGen, Google Sign-In (OAuth + PKCE). - Backend: FastAPI (Python 3.12, uv-manahe async Motor driver, deployed onRailway. - AI (OpenAI): gpt-4o for meal planning, and product matching; gpt-5.4-nano forfast conversational replies and the agentic tool-calling loop; tts-1 for spoken replies (streamed in real time for the realtime voice pilot); on-d with cloud fallback, plus an OpenAIRealtime session used purely for streaming transcription and turn-detection in the new voice pilot — reasoning is deliberately kept out of th
• Agentic layer: an OpenAI tool-calling loop that wraps existing capabilities (plan, swap, manually edit, pantry, order, log-what-you-actually-atetools Tuesday can chain.
• Context engineering: a structured user-context store plus verbatim capture of everything the user says, a product-preference memory that reasons aty vs. quality-sensitive items), and apersisted conversation transcript now surfaced directly in-app.
• Grocery automation: because Instacart /API or MCP access, we built a Chrome/Edge extension that drives the user's own logged-in Instacart session — never touching their password — with adaptive page handling and multi-tier stbrowser path (Browserbase) for a fullyhands-off cart fill.
• Feedback loop: beta testers can recordrecording-with-narration the instant theynotice something, and send it straight to us over WhatsApp — no app-store review cycle needed to hear what's broken.
• We also evaluated Google Lyria (RealTime + Lyria 3) for the music feature.

Challenges we ran into

No sanctioned way to automate groceries. The obvious path was an official integration — but Instacart isn'tapproving developer applications, and thCP, so there was no supported API or agent channel to build against. Rather than drop the feature, we found a hacky alternative: a browser extension that runs in the user's own logged-in Inutomates the cart, plus a cloud-browserfallback for when the laptop isn't around. That kept the automation and the privacy.
Local vs. deployed data split. Testing against a local backend and a deployed one with different databases caused genuine confusion about where datng, more than once, features that workedperfectly against local Mongo but silently 404'd against the deployed backend because a deploy hadn't gone out yet. We've since made deploying a roather than an afterthought. Voice in the Simulator. On-device STT anhe iPhone Simulator (and ReplayKit screenrecording doesn't work there either) — every voice-dependent feature needs a real device to actually verifywhich shaped how we test. Voice reliability is a deep well, not a ction, on-device transcription of uncommon brand/product names, and conversation memory across turns each turned out to be their own separate problem — fixing one didn't fix the others, and with regional product names, or anidle-timeout that only resets once "silence" is correctly recognized) are still open.

Accomplishments that we're proud of

• Made the assistant genuinely agentic — it takes multiple real actions, not just chat. - Real grocery automation into a live caough the user's own browser — despitethere being no sanctioned API to build on. - A preference memory that's smart aboutstaple wherever it's cheap, but the freshstuff from your trusted store. - Collapsed two confusing lists (shoppin editable order that's the single sourceof truth — and did the same for meals: swap via AI or edit by hand, your call. - A context-engineering foundation that atim and structured — for continuouspersonalization, now with a real, browsable chat history to show for it. - On-device-first voice with a graceful pilot toward continuous, interruptibleconversation. - Built a genuine, opt-in feedback loop ore ever touching the App Store reviewprocess.
  My sister — a new mom with a 22-month-old toddler — and a friend of mine both showed strong interest in trying this out, since it's built directpoints. They're now the first two betatesters.
  What we learned
  
• Personalization is only as good as the context you capture and actually use — storing raw input, not jparsed answers, matters.
• Reliably automating someone else's website is 90% timing and resilience, 10% happy path. - Tiny ordering/config bugs (how and wheend a device is actually pointed at) canmasquerade as terrifying data loss. - Scope discipline pays off — we kept thre isolated so it could never destabilizethe core, and kept the new realtime voice pilot additive rather than ripping out what already worked. - Consent has to be a first-class designught — a notification feature that'stechnically correct but not something the user explicitly opted into is still the wrong feature. - Prompt design is UX: generic prompts f on the family's real world makes theassistant feel like a collaborator.
  What's next for Coo
  
• Ship the scheduling agent (Sunday) — the second automation pillar — and capture how caretakers currentplan during onboarding so it fits their
• Fully hands-off groceries: a durable job queue that survives the laptop being offline, building furthethe cloud-hosted browser path already stInstacart API isn't available to us.
• Proactive, not reactive: extend the meal check-in pattern (opt-in, personalized timing) to more of theeveryday workflow, not just meals.
• Close the loop on voice reliability: thread real conversation memory across turns, and improve transcription accuracy for the product/b actually use.
• Track where every pantry item came from (photo scan vs. voice vs. manual) so the assistant can reason about its own confidence, not just the d
• Opt-in voice storage for research with privacy-by-design (strip PII, anonymize).
• Grow Sing & Play into record-and-remix sounds.

Built With

• adk
• codex
• gcp
• gemini
• gpt-4o
• gpt-5.6
• whisperkit