My friend reads tarot at cafes in Mumbai: Blondie in Khar on Saturdays, Janu in Juhu on Sundays, Blondie House when they have a table free. He's good at it. What he's bad at is everything around it, and so is everyone else who works from a cafe table.
Three things go wrong, and none of them are about the readings:
1.Bookings live in WhatsApp chats. Two people ask for the same 6 PM slot, and someone gets disappointed.
2.Regulars drift away. A thoughtful message after every reading would bring them back, but after six hours at the table, nobody writes twenty of them.
3.He's guessing where to sit. Saturday at Khar feels better than Sunday at Juhu. He's never checked.
So I built Walk-In.
What I Built
Walk-In is a booking platform for independent creators who work from cafés, including tarot readers, mehendi artists, caricaturists, nail artists, and musicians.
For customers: Discover nearby creators, browse services and available slots, and book without an account or online payment. Pay at the table and manage bookings through a private link.
For creators: Get a booking page, QR code, flexible scheduling, booking management, and AI-assisted messages in English, Hindi, or Hinglish.
Walk-In uses two locally hosted AI models: Gemma 4 for personalized messages and TabPFN to predict the best days, times, and cafés based on past sittings, keeping client notes private.
Real feedback: Five freelancers tested Walk-In, four said they'd pay for it, and their suggestions directly influenced the app's development.
Demo
(demo video) https://youtu.be/FMEW_k7tLPo
Live: walkin.tailf0269d.ts.net
Creators tagged Demo are dummy profiles you can safely book.
"It runs on my laptop, so if the link is down, the video shows the whole flow."
How it works
A customer opens a creator's page from the homepage or the QR code on the table, picks a service and a green time, and gets a clear "what happens next". Then the creator sees it:
Code
Walk-In
Live: walkin.tailf0269d.ts.net · Evaluation: eval/EVAL.md · Hand review: eval/REVIEW.md · A tarot reader's review: eval/CREATOR_REVIEW.md
A booking platform for creators who work from a cafe table: tarot readers, mehendi artists caricaturists, palm readers, nail artists and musicians.
Built for the Hacktoberfest Open-Source AI Challenge (Week 2: Side Hustle), around a friend who reads tarot at three cafes in Mumbai and ran his bookings out of WhatsApp chats and memory.
What it does
For customers
- Find creators by city, area, service and day ("Today", "This weekend")
- See a creator's services, prices, and where they're sitting right now
- Book a real free time in three steps, with no app needed
- Check or cancel a booking any time; an optional account keeps all bookings in one place
- Pay the creator at the session. Walk-In never handles money
For creators
- Their own booking page, with a QR code for their table
- A schedule…
How I Built It
- Backend: FastAPI, Jinja2, and vanilla JavaScript. No build step, lightweight on mobile, and Android PWA support.
- Database: SQLite with creator-specific data, separate account types, and daily backups.
- AI: Gemma 4 via Ollama and TabPFN v2 on CPU, both running locally on the same machine.
1. Best Slots Prediction — TabPFN
Problem: Creators need to know the best day, time, and café to maximize bookings and earnings.
Solution: TabPFN works well with small datasets, making it suitable for creators with just 15–50 past sittings.
Walk-In uses two regression models to predict:
- Expected earnings
- Expected customer count
It analyzes past sitting records and evaluates upcoming day × time × café combinations, recommending the best slot for each day.
Key advantage: TabPFN uses pretrained knowledge to make predictions without requiring a separate training run for every creator.
Is it better than the creator's own rule of thumb? evaluate_slots.py walks forward through the log. Each sitting is predicted from only the sittings before it, then compared with what really happened:
TabPFN — Performance Results
- More accurate: TabPFN outperformed simple averages on 4 of 6 predictions, with approximately 7% lower error and less than half the error of an overall mean.
- Fast predictions: Around 0.5 seconds per prediction on a laptop CPU.
- Best result: On a high-earning Saturday (₹17,771), TabPFN's error was ₹3,600, compared with ₹4,400 for the averages.
Limitation: The test used only 14 sample sittings, with 6 held out for prediction. The cafés are real, but the earnings data is synthetic. These results are promising, not proof of superior performance. The backtest updates as creators log real sittings.
2. Personalized, Privacy-Aware Messages — Gemma 4
After a session, creators enter brief notes, such as "Priya, confused about moving to Pune for work, drew the Star."
Gemma 4 converts these notes into a 2–4 sentence WhatsApp message, matching the creator's writing style using their previous messages.
Key features:
- Personalized: Matches the creator's tone.
- Privacy-focused: Avoids revealing sensitive session details.
- Human-controlled: Creators review, edit, and send messages themselves.
Nothing is sent automatically.
Privacy & Safety Checks — Gemma 4
Tarot sessions often involve sensitive personal information. Gemma should write thoughtful follow-ups without revealing private details.
Every draft passes through guard.py, which checks:
- Length: 2–4 sentences.
- Emojis: Maximum one; no hashtags.
- Accuracy: No invented prices or appointment times.
- Privacy: No copying 4+ consecutive words from session notes.
- Language: Correct English, Hindi (Devanagari), or Hinglish.
If a draft fails, Gemma gets one retry with specific feedback. If it still fails, the creator receives a clear warning.
Evaluation — Gemma 4 E4B
evaluate_messages.py tests 10 real-world scenarios across 3 languages and 4 prompt versions, covering thank-you messages, confirmations, follow-ups, and payment reminders.
Goal: Generate messages that are personal, accurate, private, and ready for human review.
What the Numbers Say
- Without rules: Gemma produced overly long, emoji-heavy messages, adding hashtags to 28/30 drafts. It exposed private notes twice, invented details, and failed 15/30 language checks. Average time: 46 seconds per draft.
- With clear rules: Adding limits on length, emojis, privacy, and invented details improved results to 27/30, reducing generation time to 11 seconds.
- Personalized messages: Using creators' past messages made drafts sound more natural but slightly increased Hinglish language errors, from 2 to 4.
- With a checker: One targeted retry improved results to 29/30, adding only 2 seconds. The remaining failure was flagged for manual review.
Key Problems Found
- Wrong format: 29/30 initial responses offered multiple options instead of one ready-to-send message.
- Privacy risk: One draft revealed a client's private relationship concerns.
- Hallucination: Gemma invented an appointment time of 1 PM.
- Language mismatch: Hinglish messages sometimes came out entirely in English.
- Strict validation: Some acceptable messages were flagged, but protecting client privacy mattered more than avoiding false alarms.
Key Takeaways
Clear prompts + automatic validation = more reliable AI messages.
Disabling unnecessary reasoning ("think": false) improves speed, while the checker prevents mistakes from passing silently.
Finally, 15 drafts across English, Hindi, and Hinglish are prepared in REVIEW.md for human evaluation of tone, usefulness, and respectfulness.
3. Trying to break it
An AI demo that can't take a booking is just a toy, so I tried to break the boring parts:
Why Does Open Innovation Matter?
1. Privacy First
Creators handle sensitive client information. Running Gemma 4 locally through Ollama keeps private session notes on a trusted machine instead of sending them to third-party AI APIs.
2. Lower Costs
Local AI eliminates per-token API charges, making personalized follow-ups affordable for small creators. TabPFN also turns limited sitting records into useful predictions without requiring large datasets.
3. Freedom to Customize
Open models allowed me to:
- Disable unnecessary reasoning for faster responses.
- Run Gemma on GPU and TabPFN on CPU.
- Test prompt improvements against 30 evaluation cases.
4. Open vs. Closed Models
Closed models might produce smoother Hindi and need fewer retries, but they still require privacy safeguards and output validation.
Key takeaway: Open innovation makes Walk-In more private, affordable, customizable, and accessible for independent creators.
My First Users
Walk-In was tested by 5 freelancers: a tarot reader, musician, photographer, tutor, and designer. 3 tried the app, while 2 reviewed its features.
Key Feedback
- 4 out of 5 said they'd pay at least ₹99/month.
- Users liked the QR code, booking system, and WhatsApp integration.
- The biggest criticism was the empty marketplace and unclear user journeys.
Improvements Made
- Demo creators added across three cities, with sample services and bookings.
- Separate customer and creator interfaces for easier navigation.
- Better onboarding with help pages and feature explanations.
- New filters for city, area, and availability.
- Private booking management for checking or cancelling appointments.
A Tarot Reader's Review
"Usually, I have to explain my services, prices, and availability to every person who messages me. Having everything on one booking page would make things much easier."
They also appreciated the QR code and AI-assisted follow-ups, especially the ability to edit messages before sending.
What's Next?
- Interactive map to discover nearby experiences.
- One-click rescheduling instead of cancelling and rebooking.
Key takeaway: Although the sample size was small, feedback from real freelancers helped identify problems and directly shaped Walk-In's improvements.
Prize Categories
•Gemma: Gemma 4, run locally through Ollama, drafts every client message in English, Hinglish or Hindi in the creator's own voice. Every draft goes through a rule checker with one retry, and four prompt versions are evaluated on the same 30 cases.
•TabPFN: TabPFN v2 predicts earnings and customer counts from each creator's small log, ranks next week's best day, time and cafe, and is backtested against the creator's own averages.
Top comments (0)