Loafline: Open-Weight AI That Gives You Permission to Defer (Hacktoberfest Week 3)

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady.

What I Built

Loafline answers a question most productivity apps pretend doesn't exist: "Can I safely loaf on this commitment?"

You paste text or record a short voice memo. Open-source AI returns:

• a defer verdict (loaf / this_week / today) from TabPFN on tabular features,
• the smallest closing action so you are not ghosting anyone,
• a safe-until date,
• and a markdown permission slip you can copy into Slack or email.

The Sloth ledger (/history) stores every check on MongoDB Atlas, surfaces similar past promises from Tiger Data (pgvector), and can generate a Gemma morning digest. Batch mode triages up to five standup brain-dump lines in one pass.

Lazy payoff (measured on production): our Render smoke test estimated ~14 minutes of rumination avoided per check versus manually re-listening, drafting, and worrying.

Demo

• Live app: https://loafline.onrender.com
• Health: https://loafline.onrender.com/api/health
• Suggested 60s flow: voice memo → verdict → Copy permission slip → Ledger stats → batch mode

Code

smriad / loafline

Loafline

Loafline is a Hacktoberfest Week 3 (Slow and Steady) project: open-weight AI that rewards laziness by telling you which commitments you can safely defer, with the smallest closing action and a “safe until” date.

What it does

1. You paste text or record a short voice memo.
2. Gemma (open-weight via OpenRouter) extracts who/what/when.
3. TabPFN classifies defer risk from tabular features (with a bundled training set).
4. Tinker polishes the sloth-friendly verdict copy.
5. Backboard recalls patterns; SerpApi optionally grounds deadlines.
6. Tiger Data (pgvector) surfaces similar past promises.
7. MongoDB Atlas stores the ledger; ElevenLabs transcribes voice and reads the verdict aloud.
8. Sentry traces the pipeline; Mastra orchestrates the workflow.
9. Deploy on Render (render.yaml).

Local dev

cp .env.example .env.local
# fill keys
npm install
npm run dev

Set DEMO_MODE=true to try the UI without APIs.

Render deploy

Repo: github.com/smriad/loafline

Option A — Blueprint (recommended)

1. Go to dashboard.render.com → New →…

View on GitHub

How I Built It

Loafline is a Next.js 15 app on Render (render.yaml). A Mastra workflow wraps a single pipeline step that orchestrates partner APIs server-side.

flowchart LR
  A[Voice or text] --> B[ElevenLabs STT]
  B --> C[Gemma extract]
  C --> D[TabPFN defer class]
  D --> E[Tinker verdict copy]
  E --> F[Backboard memory]
  F --> G[SerpApi ground]
  G --> H[Tiger echo search]
  H --> I[(MongoDB ledger)]
  I --> J[Permission slip + lazy payoff]

Partner	What it does in Loafline
Render	Hosts the web app and API
Gemma (open-weight via OpenRouter)	Extracts who / what / deadline from messy input
TabPFN	Classifies defer risk from feature rows (word_count, deadline words, …)
Tinker	Polishes sloth-friendly verdict prose
Mastra	Workflow orchestration for /api/loaf
MongoDB Atlas	Persistent commitment ledger
Tiger Data	pgvector echoes — "you promised this before"
ElevenLabs	Speech-to-text + verdict text-to-speech
Backboard	Cross-session memory on recurring defer patterns
SerpApi	Optional deadline / context grounding
Sentry	Traces the agent pipeline (loafline.run span)

Production partner wall currently reports 11/11 integrations live after env vars are set on Render.

Why Does Open Innovation Matter?

A closed "productivity" API would own your guilt graph: every promise, every voice memo, every pattern of avoidance. Loafline keeps the open pieces at the center:

• Gemma (open-weight) does the language understanding — you can swap models or routers.
• TabPFN scores defer risk from auditable tabular features shown in the UI, not a black-box "urgency score."
• Voice and ledger data flow through your MongoDB / Tiger services, not a vendor training set.

That matters for Slow and Steady: the win is not hustle — it is doing less mental work without handing your commitments to a single closed model vendor.

The Lazy Payoff

Scenario	Manual (est.)	Loafline
Re-listen to a voice memo + draft a reply	~8–15 min	~2.5 min (voice path)
Monday standup — 5 loose promises	~40+ min of triage	Batch mode (5 lines, one flow)

The app surfaces lazyPayoffMinutes on every verdict and aggregates minutesSavedTotal on the Ledger page — numbers you can screenshot for judges.

Prize Categories

I'm entering:

• Best Use of Render
• Best Use of Gemma
• Best Use of TabPFN
• Best Use of Tinker
• Best Use of MongoDB Atlas
• Best Use of Tiger Data
• Best Use of ElevenLabs
• Best Use of Backboard
• Best Use of SerpApi
• Best Use of Sentry Agent Tracing
• Best Use of Mastra

---

Built for Hacktoberfest 2026 — AI belongs to everyone. Questions welcome in the comments.