Chithi

A voice-first AI assistant that takes the invisible mental load of motherhood off a mom'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

Motherhood 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 mom'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 Chithi.

Chithi is the tamil word for Mom's younger sister, usually symbolising a close family member, known to be a caretaker of the mom and the kid. Cherry on the top, I am a Chithi myself to a toddler, whose caretaking journey inspired this product.

What it does

Chithi is a voice-first assistant that works as a system of agents taking care of various everyday workflows that saves a lot of Mom'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, and lets you swap any meal.
• Reads your fridge/pantry from a photo and keeps a live inventory.
• Builds a shopping list, splits it across your preferred stores, and remembers your brand and store preferences.
• Adds items to your Instacart cart through a browser helper running in your own logged-in browser.
• It's agentic — a single request can trigger several actions ("plan my week and add what's missing to my order").

The scheduling agent, Sunday, is next in line. There's also an experimental agent DJ which helps parents engage young kids and toddlers, by jamming on the favorite beats and mixing it up with their own themes, vocals and characters, this is a Sing & Play mode where you make up songs with your kid over familiar tunes, with more 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-managed), MongoDB Atlas with the async Motor driver.
• AI (OpenAI): gpt-4o for meal planning, pantry vision, reasoning, and product matching; gpt-5.4-nano for fast conversational replies and the agentic tool-calling loop; tts-1 for spoken replies; on-device Apple/WhisperKit STT with whisper-1 as a cloud fallback.
• Agentic layer: an OpenAI tool-calling loop that wraps existing capabilities (plan, swap, pantry, order, remember-preference) as tools Tuesday can chain.
• Context engineering: a structured user-context store plus verbatim capture of everything the user says, and a product-preference memory that reasons about where to buy (commodity vs. quality-sensitive items).
• Grocery automation: because Instacart no longer offers developer/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 store routing.
• 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't approving developer applications, and they've discontinued their MCP, 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 Instacart session and semi-automates the cart. That kept the automation and privacy.

Local vs. deployed data split. Testing against a local backend and a deployed one with different databases caused genuine confusion about where data actually lived, thanks to claude I could locate it before it got too late.

Voice in the simulator. On-device STT and the mic didn't work in the iphone Simulator, but could fix it again with some authentication steps, which sped up testing without the dependency for my iphone.

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 cart, done privacy-first through the user's own browser — despite there being no sanctioned API to build on.
• A preference memory that's smart about routing: buy the branded staple wherever it's cheap, but the fresh stuff from your trusted store.
• Collapsed two confusing lists (shopping list vs. order) into one editable order that's the single source of truth.
• A context-engineering foundation that captures everything — verbatim and structured — for continuous personalization.

• On-device-first voice with a graceful cloud fallback.

• My sister a new mom with a 22-month old toddler, and a friend of mine both have shown strong interest in trying out this product, as its directly built to address their biggest pain points.

What we learned

• Personalization is only as good as the context you capture and actually use — storing raw input, not just parsed answers, matters.
• Reliably automating someone else's website is 90% timing and resilience, 10% happy path.
• Tiny ordering/config bugs (how and when env loads) can masquerade as terrifying data loss.
• Scope discipline pays off — we kept the experimental music feature isolated so it could never destabilize the core.
• Prompt design is UX: generic prompts feel like a form; anchoring on the family's real world makes the assistant feel like a collaborator.

What's next for Chithi

• Ship the scheduling agent (Sunday) — the second automation pillar — and capture how moms currently plan during onboarding so it fits their real workflow.
• Fully hands-off groceries: a durable job queue that survives the laptop being offline, and a cloud-hosted browser so the automation runs without the user's machine — since a sanctioned Instacart API isn't available to us.
• Proactive, not reactive: actually act on the context and automation wishes we capture.
• Deploy (it's currently local-first).
• Opt-in voice storage for research with privacy-by-design (strip PII, anonymize).
• Grow Sing & Play into record-and-remix, built from the kid's own sounds.

Built With

• digitalocean
• fastapi
• gemini
• gpt
• lyria
• mongodb
• python
• railway
• swift