House Pot: Amma's kitchen in Dhaka — approve first, then listen

Submission for Hacktoberfest Weekend: Build for a Friend — #hf26challenge

House Pot is a small kitchen app for Amma in Dhaka: one recipe from tonight’s pantry, allergies checked in code, voice only after she approves.

The problem

Weeknight dinner in our flat is not “pick a trending recipe.” It is what came back from the market, who is eating, and who cannot eat peanuts or shellfish. Recipe apps and generic AI chefs assume you will shop for their list, trust a prompt for safety, and read long steps while the rice is on the stove. That fails when:

• the pantry is fixed for tonight (dal, rice, greens, whatever was affordable)
• allergy mistakes are not acceptable—a cousin’s peanut reaction is not a disclaimer in a chat box
• the cook wants one clear answer, not ten options and a wall of text
• voice help is welcome, but not before she has seen and accepted the exact dish

Who it's for

Amma—the person who actually runs our stove in Dhaka. She improvises from habit and memory, carries the household allergy list in her head, and does not want to feel like she is “talking to an AI chef.” House Pot is also for family members who share the same allergies and pantry, and for anyone in a similar setup: one cook, real constraints, approve-before-you-listen.

Defaults in the app match our kitchen: cook name Amma, allergies peanuts and shellfish, pantry like red lentils, onion, garlic, rice, cumin, spinach, yogurt.

How I'm solving it

Goal	Approach
One dinner, not a feed	Gemma (GEMMA_*, OpenAI-compatible) returns one structured JSON recipe; optional critic pass may revise once
Safety you can see	pantry-check.ts marks in-kitchen vs missing and blocks Approve on allergen matches (shrimp → shellfish, peanut oil → peanuts); Open Food Facts when available—not “the model said it’s fine”
No surprise audio	ElevenLabs runs only after Approve & read aloud on that card; approve and narrate re-run the same checks (idempotent narrate on retry)
Memory across nights	MongoDB Atlas (or local .data/house-pot.json fallback) stores household, pantry habits, run history, and feedback (“less cumin next time”)
Hands busy at the stove	Whisper (local STT) for pantry dictation; narration reads approved recipe text only, never the raw voice note

Design split: plan (probabilistic Gemma JSON) → policy (deterministic pantry + allergen code) → deliver (TTS gated on approve). Human approval binds to this card version; a stale “yes” cannot narrate a recipe that now fails checks.

Requirement (Amma)	Engineering
Cook from tonight’s pantry	POST /api/runs → orchestrator propose
Allergies enforced outside the model	PantryReview.canApprove / safeToNarrate
Auditable for family / judges	Ingredient marks on card; GET /api/runs/:id/trace

Full BRD table, C4 diagrams, and state machine: README on GitHub.

Real run (live smoke): pantry = red lentils, onion, garlic, rice, cumin, spinach, yogurt; allergies = peanuts and shellfish; three diners. Gemma proposed Mild Spinach and Red Lentil Dal with Rice. Every line matched the pantry; narration waited until approve.

Architecture

Stack: Next.js on Render (demo) · Ollama + Gemma (local plan) · MongoDB Atlas · pantry-check · ElevenLabs TTS after approve · Mastra approval gate · Temporal narration when the worker is up.

sequenceDiagram
  participant U as Browser
  participant API as Next.js API
  participant O as orchestrator
  participant G as Gemma
  participant P as pantry-check
  participant E as ElevenLabs

  U->>API: POST /api/runs
  API->>O: startKitchenRun
  O->>G: propose JSON recipe
  O->>P: reviewPantry(allergies)
  P-->>O: canApprove + marks
  API-->>U: recipe card

  U->>API: POST approve
  API->>P: re-check
  U->>API: POST narrate
  API->>P: re-check
  API->>E: TTS approved text only

Propose path (simplified): load household memory → Gemma proposes recipe (optional critic) → deterministic pantry + allergen review → show card with marks → cook taps approve → ElevenLabs narrates that text.

Local vs live: same codebase. My machine runs Ollama gemma3:4b, Whisper, TabPFN meal-fit, Temporal + Mastra, Backboard + Tiger memory mirrors, Sentry traces—see GET /api/health when npm run dev is up. Live demo uses hosted GEMMA_* on Render so judges click without Ollama; TABPFN_DISABLE=true there keeps propose fast on free tier.

Demo

Live: https://house-pot.onrender.com/

Demo video (~5 min): https://house-pot.onrender.com/demo.mp4 — 1080p UI tour (male Bangladesh-style voice); each chapter names which technology, where in the UI/API, and how it is used, then recipe TTS (npm run demo:record)

Example run (pre-publish pass): https://house-pot.onrender.com/?run=bc1c4411-5d4b-42a3-ab04-97b0346b53ea — Mild Spinach and Red Lentil Dal with Rice, narrated

Example run history (smoke test): https://house-pot.onrender.com/history?household=8c596887-5372-4c81-bd95-8ec1babe627b

No login—the browser saves one household id on first visit (or open history with ?household=<uuid>).

Path for judges (Render free tier: first load after sleep can take ~60 seconds):

1. Open the link; wait for the kitchen screen.
2. Cook name Amma; allergies peanuts, shellfish.
3. Pantry example: red lentils, onion, garlic, rice, cumin, spinach, yogurt.
4. Tap Propose tonight's pot; read which items are marked in-kitchen vs missing.
5. Tap Approve & read aloud only when the card is safe; audio runs after that gate.

If approve is disabled, the card shows why—a missing market item or an allergen. That is the product.

Skeptics: src/lib/kitchen/pantry-check.ts owns the marks; approve and narrate re-run the same logic so an old “yes” cannot read aloud a recipe that now includes shrimp.

Code

https://github.com/smriad/house-pot

Path	Role
src/lib/gemma.ts	OpenAI-compatible JSON recipe; extractRecipeJsonText for messy model output
src/lib/kitchen/pantry-check.ts	Pantry match + allergy block (system of record)
src/lib/kitchen/orchestrator.ts	Propose → check → approve → narrate
src/lib/kitchen/critic.ts	Optional second Gemma pass
src/lib/db/store.ts	MongoDB Atlas + .data JSON fallback
src/lib/mastra/kitchen-workflow.ts	Approval suspend when Mastra storage is up
src/components/HousePotApp.tsx	Kitchen UI
render.yaml	Render Blueprint (TABPFN_DISABLE, TEMPORAL_NARRATE=false)

Technologies — what I used, how, and why

I only list tools that are live on my machine when I run npm run dev (GET /api/health). I did not wire SerpApi (no key). Repo stubs for other sponsors exist for probes, but they are not part of my daily path.

Production check (/api/health on Render): Gemma, MongoDB, ElevenLabs, Mastra, Backboard, Tiger, Sentry, TheMealDB, and Open Food Facts are live; Whisper, Temporal, embeddings, TabPFN, and SerpApi are off there by design (Node-only blueprint, TABPFN_DISABLE, no Temporal worker).

Core application

Technology	How I used it	Why I used it
Next.js 16 (App Router)	Kitchen UI + /api/runs, approve, narrate, health	One app for Amma’s screen and server-side safety checks
React 19 + TypeScript	Recipe card, approve gate, run history	UI matches orchestrator states (awaiting_approval → narrated)
Tailwind CSS 4	Touch-friendly kitchen layout	Fast iteration at the stove
Zod	recipeSchema on every Gemma response	Catch bad JSON before it hits the card
OpenAI Node SDK	gemma.ts → GEMMA_* (Ollama locally, hosted on Render)	One client for local open weights and the public demo

Planning (probabilistic)

Technology	How I used it	Why I used it
Ollama + Gemma gemma3:4b	proposeRecipe() → one JSON recipe from pantry + allergies	Open-weight planning on my laptop
Gemma critic	Second pass may replace the draft once	Better card before Amma reads it; safety still in code
Cook brief	One-line summary on the card	Less reading while rice cooks
Embeddings (nomic-embed-text via Ollama)	Rank pantry memories on propose	“Less cumin” surfaces without retyping the story
Hosted GEMMA_* (Render only)	Same code path on https://house-pot.onrender.com/	Judges try the flow without Ollama

Policy (deterministic — the product)

Technology	How I used it	Why I used it
pantry-check.ts	In-kitchen / missing marks; blocks approve on allergens	Allergy control is code, not a prompt
Re-check on approve + narrate	Same review on both routes	Stale “yes” cannot read aloud a new allergen
Open Food Facts	Allergen tags when enriching ingredients	Extra signal beyond string match
TheMealDB	Public dish names on propose	Ground the planner without another API key

Voice

Technology	How I used it	Why I used it
ElevenLabs	TTS only after approve; idempotent narrate	Clear steps for Amma—never the pantry voice note
Whisper (local Python)	POST /api/transcribe for pantry dictation	Hands busy; STT stays on my machine

Memory and persistence

Technology	How I used it	Why I used it
MongoDB Atlas	Households, runs, memories, feedback	Tuesday’s lentils are not a new lecture
Backboard	Semantic cook memories on propose / feedback	Longer-thread kitchen context
Tiger Data (Postgres)	Mirrors feedback after a run	SQL sidecar for memory experiments
localStorage	Browser household id	No login; share history with ?household=

Workflows, scoring, observability

Technology	How I used it	Why I used it
Mastra + LibSQL	Approval gate suspends until she taps approve	Human-in-the-loop as a workflow step
Temporal	Durable narrate when worker + TEMPORAL_ADDRESS are up	Safe retries locally (TEMPORAL_NARRATE=false on Render)
TabPFN (Python)	Blended with heuristic friend-fit score on propose	“Would Amma cook this?” signal in dev
Sentry	Spans on agent steps	Trace propose → approve → narrate
Run trace	GET /api/runs/:id/trace	Show judges each step

Demo and deploy

Technology	How I used it	Why I used it
Render	render.yaml → live URL for the challenge	Public demo
GitHub Actions	live-smoke.yml	CI hits propose on production
Playwright	npm run demo:record → demo.mp4	Video for judges who skip cold start

Design takeaway

Plan (Gemma on Ollama) → policy (pantry-check) → deliver (ElevenLabs, gated). Memory and workflow tools support that line—they do not replace it.

Human-in-the-loop only works when the card shows what is on the shelves and what would be unsafe—not a blind “trust me” approve button.

Why Does Open Innovation Matter?

In a Dhaka household the sensitive data is not “inspiration.” It is who reacts to peanuts, what Amma actually bought before the monsoon rain, and what we cannot send to a closed meal API abroad.

Open-weight Gemma can plan on a machine we run—laptop, small server, or a GPU droplet if we outgrow Ollama. The validation logic stays ours. The Render link uses hosted inference so judges anywhere can click; I would rather say that plainly than pretend the demo box is offline-only.

ElevenLabs is the closed piece, and it sits behind approve—it never sees the voice note, only the recipe she accepted.

MongoDB holds what she would otherwise repeat at every family lunch: allergies, pantry habits, “a little less chili next time.”

Prize Categories

• Best Use of Gemma — dinner from the pantry + constraints; structured JSON recipe.
• Best Use of ElevenLabs — narration only after approve; not before.
• Best Use of MongoDB Atlas — household memory across nights in Dhaka and on the live URL.
• Best Use of Render — https://house-pot.onrender.com/ from render.yaml.

My Agent Session

Curated build log: Playwright demo.mp4 pipeline, per-chapter tech voiceover (technology / where / how), Bangladesh-style narrator, and syncing this submission draft to DEV.

House Pot — demo video, tech voiceover, and DEV submission sync

You

Prepare House Pot for Hacktoberfest Weekend submission: live demo on Render, DEV draft, and a judge-friendly demo video.

Agent

Expanded README/SUBMISSION with BRD traceability, plan→policy→deliver architecture, and honest tech tables tied to GET /api/health. Fixed Render propose latency (TABPFN_DISABLE), Mastra probe, and smoke-test run IDs for judges.

You

Record demo.mp4 locally with visible cursor and voice that explains each feature.

Agent

Added npm run demo:record: Playwright 1920×1080 capture, chapter voiceovers via ElevenLabs, ffmpeg mux. Scripts live under scripts/record-demo.mjs and scripts/lib/demo-voice.mjs.

You

Narration should cover which technology is used, where in the UI/API, and how it is used — Bangladesh male accent.

Agent

Rewrote tour chapters (Next.js, health strip, MongoDB profile, STT input, Gemma orchestrator, pantry-check policy, ElevenLabs gate, run history, recipe TTS). Default demo voice: Aashish (South Asian male). Re-recorded public/demo.mp4 (~5 min).

You

Update the DEV submission draft and agent session embed.

7 of 7 messages

Session on DEV: House Pot — demo video, tech voiceover, and submission sync

Friend quote

“The dal was right—use a little less cumin next time. And I will not listen to the voice until I tap approve.”

— Amma, after the first real dinner from House Pot.

---

Built for Hacktoberfest Weekend 2026 — Build for a Friend, from a Bangladesh kitchen.