Tangkwa Journal: A Soft AI Diary I Built for My Friend in ENT Training

Mochamad Faishal Amir — Sat, 03 Oct 2026 06:46:02 +0000

This is a submission for the Hacktoberfest Weekend Challenge: Build for a Friend

What I Built

I built Tangkwa Journal for my friend Tangkwa.

She’s in a demanding ENT specialist training program. Her days are long, her schedule is messy, and when she’s rushed or stressed she often reaches for something sweet — a cup of sweet matcha, a quick comfort drink — without really tracking how often it happens. She doesn’t need another calorie spreadsheet. She needs something that feels like a friend.

So I made a cute, mobile-first PWA where she can:

Write a short diary entry (optional mood chips: tired, okay, stressed, happy, overwhelmed)
Log a meal or drink in plain language — “a cup of sweet matcha”, “a slice of pizza”
Get a warm reply from an open-weight companion that acts as a psychologist-style friend and a gentle diet mentor

The brain is gemma2:2b via Ollama. The phone is a thin PWA. Inference stays on the server side on purpose — her notes stay on an open-weight stack I control, not a closed chatbot SaaS in the middle.

Accounts are per-person (username + password, httpOnly session cookie) so each journal stays private. Important: the app is explicit that this is a supportive friend, not a clinician. There’s an in-app disclaimer. Crisis-looking language takes a fixed human-help path and skips model generation.

Demo

Live demo: https://tangkwa-journal.byfaishal.cv/

Judges can Create an account on the live site, or sign in with the seed account:

username: tangkwa
password: tangkwa-journal

Local / judge repro:

git clone https://github.com/mochfamir/tangkwa-journal
cd tangkwa-journal
pnpm install
ollama pull gemma2:2b
cp .env.example .env
pnpm dev:api   # http://127.0.0.1:8787
pnpm dev:web   # http://127.0.0.1:5173

Health check (must see the open-weight model):

curl -s http://127.0.0.1:8787/api/health
# {"ok":true,"model":"gemma2:2b","ollama":"up","modelPresent":true}

Companion smoke test (register → cookie → create meal entry):

curl -s -c /tmp/tj.jar -X POST http://127.0.0.1:8787/api/auth/register \
  -H 'content-type: application/json' \
  -d '{"username":"judge-demo","password":"judge-demo-pass","display_name":"Judge"}'

curl -s -b /tmp/tj.jar -X POST http://127.0.0.1:8787/api/entries \
  -H 'content-type: application/json' \
  -d '{"type":"meal","body":"a cup of sweet matcha","sweetness":"sweet"}'

Or skip register and log in as the seed user (tangkwa / tangkwa-journal) against /api/auth/login with the same cookie jar.

Code

https://github.com/mochfamir/tangkwa-journal

Built during the challenge window (commits from 2026-10-03, after the Oct 2 UTC start).

How I Built It

I shipped in deliberate layers the same Saturday afternoon (WIB), because the challenge is short and Tangkwa’s problem is real.

1. Brain first — open-weight companion loop

Before pastels, I wanted proof that Gemma could talk back.

Small Hono API (apps/api) on port 8787
SQLite (better-sqlite3) for diary + meal entries, scoped per user
Versioned system prompt in prompts/companion.md — friend first, coach second, ~100-word cap, no shame language, no fake calorie precision
Ollama POST /api/chat with gemma2:2b and a tight num_predict
Only the last 3 one-line summaries as context (keeps a 2B model coherent; avoids dumping her whole journal into the prompt)
Fail closed: if Ollama is down → HTTP 503, nothing half-written is persisted

Warm replies on my machine landed in about 1–4 seconds. That was the moment it stopped being a scaffold and became something that could answer a tired ENT trainee who just logged a sweet matcha.

Sample meal reply from build-time testing:

Matcha's a great pick, Tangkwa! Sometimes a little sweetness helps bridge the day's fatigue. How's that energy after rounds?

2. Then the shell she’d actually open

UI only — no API contract changes, so the open-weight path stayed honest.

Vite + React + TypeScript PWA
Soft pastel system (blush / cream / mint) with Fraunces + Nunito
Screens: login/register, Today (“Hi Tangkwa”), Write diary, Log drink/meal, History, Detail with AI reply, About/disclaimer
“Gemma is thinking…” loading state
PWA shell via vite-plugin-pwa — offline shell is fine; /api/* stays network-only so a dead model never becomes a fake journal reply in the UI

Architecture in one line:

Phone PWA  →  Hono API (+ session auth)  →  Ollama (gemma2:2b)
                 ↓
              SQLite journal (per user)

3. Why `2b`, not a bigger Gemma?

I already had a larger Gemma locally. I still chose gemma2:2b so a weekend deploy on modest hardware stays realistic, latency stays friendly on a phone wait screen, and the “live for a friend” path doesn’t depend on a GPU fairy tale. Quality is managed with a strict prompt and short outputs — not by hoping a bigger model forgives a vague product.

Privacy-minded journal builders keep saying the same thing: personal notes need a hard boundary on where inference runs and what gets remembered. That pushed me toward self-hosted open weights instead of a closed chat API as the product’s brain.

Why Does Open Innovation Matter?

This isn’t a todo list. It’s mood + eating notes from someone I care about, written on exhausted hospital days.

I didn’t want that intimacy routed through a closed commercial chatbot by default.

Open-weight Gemma + open-source Ollama means:

The model is a setting, not the product’s landlord — swap weights later without rewriting around a vendor
Inference can stay on infrastructure I run (laptop now; VPS/Compose for her phone later)
Judges (and Tangkwa) can see the companion voice in a file in the repo, not buried in a black-box prompt dashboard
A small open model is enough for short, kind check-ins — which is what she actually has time for

Closed APIs would have been faster to prototype. They would have been the wrong default for this friend’s data.

Prize Categories

Best Use of Gemma — generation path is only gemma2:2b via Ollama; no closed-LLM fallback in the companion loop. The PWA is a thin client; Gemma is the product’s brain.

Built for

Tangkwa — and anyone who’s too busy for calorie apps but still deserves a soft place to check in.

DEV Community: Mochamad Faishal Amir