Chhath Ghat Guide: an offline AI helper for the day the network dies at the ghat

The ghat, the crowd and the dead network Every year during Chhath Puja, lakhs of families walk down to the ghats of Bihar to offer arghya to the setting and the rising sun. Vratis stand waist-deep in the river, children hold their parents' hands, elders watch the water, and the songs carry across the steps. And every year the same three things happen: Families get separated in the crowd. Notices about deep water, barricades and timings are hard to read in the press of people — and many elders are more at home in Bhojpuri or Maithili than in printed Hindi. The mobile network jams. Calls don't connect, messages don't send, and every app that needs the internet stops working — exactly when it's needed. So I built a small helper for that moment. Chhath Ghat Guide runs on one laptop at the ghat. The laptop's Wi-Fi hotspot has no internet at all; family phones join it and open the app in a browser. The AI model, the notes and every photo stay at the ghat. The screen is meant to be a few seconds of help, not the experience. The experience is the river, the sun and the family. Three tabs, one job each. (Screens are from the real app on a phone-sized browser; see the note under each image about what's test data.) Take a photo of a notice or banner. A local Gemma 3 (4B) vision model: copies the text exactly as written, and explains it in simple Hindi, Bhojpuri or Maithili — using only the text it just read. The original words are always on screen. Safety notices (deep water, barricades, emergency) show the original text first, inside a red box — never only a summary. A blurry photo gets "साफ़ नहीं पढ़ पाए" ("couldn't read it clearly") and tips for a better photo, instead of a guess. And if the explanation contains a number that isn't in the notice — an invented ghat number or time — the app warns you. Real app screens; the notices are printed samples, and the read text and explanations shown are test data copied from the notices (screens with Gemma's own output will replace these after the field test). Ask "सूप में क्या रखते हैं?" (what goes in the soop?) or "खरना में क्या बनता है?". Answers come only from ten short notes written by the family and checked by an elder — not from the model's memory. Each answer: shows which note it came from and who checked it, ends with a reminder that customs differ between families and regions, and falls back to "यह मेरे नोट्स में नहीं है — किसी बड़े से पूछें" ("not in my notes — ask an elder") when the notes don't cover it. Questions about clock times, this year's dates or parking never reach the model at all — they're sent to official notices and the panchang. An AI should not be the one telling a vrati what time to stand in the river. Real app screens; the answer text is test data copied from the note it cites. The notes are marked "डेमो — समीक्षा बाकी" (demo — review pending) because the elder review comes before the real demo. "Save our spot" stores the GPS position, a photo of a landmark and a note ("नीला टेंट, घाट 5 की सीढ़ियों के पास" — blue tent by the ghat 5 steps) on the phone only. "Take me back" shows the distance and an arrow that turns with the phone's compass. GPS comes from satellites, not the mobile network, so this keeps working when calls don't — and because it runs entirely in the phone's browser, it keeps working even if the phone wanders out of the laptop's Wi-Fi. When you're within GPS error of the spot, the arrow would just be noise, so it switches to "आप बहुत पास हैं — फोटो वाली निशानी देखें" (you're very close — look for the landmark in the photo). The spot deletes itself at midnight. Real app screens with emulated GPS and compass (the landmark is a drawing). An emergency card sits on every screen and sends people to the nearest police or volunteer booth. The app ships with no phone numbers — organisers add verified ones on the day; a made-up number at a crowded river is worse than none. Part Choice Why Model Gemma 3 4B via Ollama, for both reading photos and writing answers Open weights, runs on a laptop CPU, no internet App Python + Streamlit, three tabs Fast to build; camera and file inputs built in Ritual answers 10 Markdown notes, keyword + nomic-embed-text search Answers traceable to checked text Meeting Spot Plain HTML + JS, Geolocation + DeviceOrientation APIs Runs on the phone; location never leaves it Phones Self-signed HTTPS on the hotspot Phone browsers only allow GPS, camera and compass on secure pages The interesting part of this project wasn't getting Gemma to talk — it was making sure it doesn't say more than it knows, in a place where a wrong answer can matter. The Notice Reader makes two calls on purpose. The first copies the text (temperature 0, unreadable words marked [?]); the second explains only that text, so every claim on screen can be checked against the original right below it. # ghat_guide/notice_reader.py (trimmed) def transcribe(self, image: bytes) -> Transcription: out = self.llm.chat_json( model=self.models.vision_model, system=prompts.SYS_TRANSCRIBE, # "copy exactly; write [?] for unreadable words" user=prompts.USER_TRANSCRIBE, schema=prompts.TRANSCRIBE_SCHEMA, # {text_read, script, clarity} temperature=0.0, images=[image], ) clarity = clarity_verdict(out["text_read"], out["clarity"], ...) # model's opinion + our own rules return Transcription(out["text_read"], ..., clarity, is_safety_notice(out["text_read"]), ...) def explain(self, tr: Transcription, lang: str) -> Explanation | None: if tr.clarity == "unclear": return None # nothing reliable to explain: show "साफ़ नहीं पढ़ पाए" out = self.llm.chat_json(model=self.models.text_model, system=prompts.sys_explain(lang), user=prompts.user_explain(tr.text_read), ...) return Explanation(..., number_warning=unsupported_numbers(out["explanation"], tr.text_read)) Splitting the steps had a nice side effect: switching language re-runs only the cheap text step, not the slow image step. The safety check is plain keyword matching, chosen carefully — "डूब" (drown) alone would also match "डूबते सूर्य", the setting sun that Sandhya Arghya is offered to, so the list uses "डूबने" instead: # ghat_guide/safety.py (excerpt) # "डूब" alone would match "डूबते सूर्य" (the setting sun of Sandhya Arghya), # "मना" alone would match "मनाया" (celebrated). SAFETY_TERMS_DEVANAGARI = ("गहरा", "गहराई", "डूबने", "खतरा", "बैरिकेड", "निषेध", "मना है", ...) # ghat_guide/ritual/answerer.py (trimmed) def ask(self, question: str, lang: str) -> RitualAnswer: if asks_for_official_info(question): # "बजे", "तारीख", "इस साल", "पार्किंग", ... return RitualAnswer("official_info", None) # never asks the model relevant = [h for h in self.retriever.search(question) if self.retriever.is_relevant(h)] if not relevant: return RitualAnswer("not_in_notes", None) out = self.llm.chat_json(..., user=prompts.user_ritual(context, question), schema=prompts.RITUAL_SCHEMA) # {found, answer, used_chunk_ids} given = {h.chunk.chunk_id for h in relevant} cited = out.get("used_chunk_ids") or [] if not out.get("found") or not cited or any(c not in given for c in cited): return RitualAnswer("not_in_notes", None) # no sources, or sources it wasn't given return RitualAnswer("answered", out["answer"], sources=[...]) Search is a mix of keywords and embeddings, because nomic-embed-text is mostly trained on English and Devanagari questions alone can rank poorly. Hand-written keywords plus a small synonym map (अरघ → अर्घ्य, सूपा → सूप) and a light stemmer (खरने ↔ खरना) catch the words families actually use. And a note without an elder's checked_by is never loaded at all. The Meeting Spot runs in the phone's browser, so it's plain JavaScript: // static/meeting_spot.html (excerpt, simplified) function bearing(lat1, lon1, lat2, lon2) { // degrees clockwise from north var p1 = lat1 * rad, p2 = lat2 * rad, dl = (lon2 - lon1) * rad; var x = Math.sin(dl) * Math.cos(p2); var y = Math.cos(p1) * Math.sin(p2) - Math.sin(p1) * Math.cos(p2) * Math.cos(dl); return (Math.atan2(x, y) / rad + 360) % 360; } // heading comes from the compass (iOS: webkitCompassHeading, Android: 360 - alpha), // smoothed with a circular mean so 350° and 10° average to 0°, not 180°. arrow.style.transform = "rotate(" + mod360(bearing(here, spot) - heading) + "deg)"; The same maths lives in Python (ghat_guide/geo.py), and a test runs both side by side in a real browser so they can't drift apart. A real run of scripts/check_offline.py — the warnings are true: the notes are still waiting for the elder review. It checks that the models are downloaded, the ritual index is built, the phone pages load nothing from the internet, the HTTPS certificate matches the hotspot's IP (a classic "it worked at home" failure), and that the model server isn't reachable from the phones. [FILL — after the field test. Suggested shape: Where: the ghat or riverbank, the date, who came along. The moment you switched mobile data off and the app kept working. Numbers from docs/FIELD_TEST.md: seconds per notice, whether the arrow pointed the right way, GPS accuracy near the water. What the elder said about the Ritual Helper (with permission), and what they corrected in the notes. What didn't work — honestly.] The real test is still ahead: Chhath Puja is in late October / early November, and I'll take it to the ghat then and report back here. Why open innovation matters here It works with no network. A cloud AI is useless when the network is jammed; an open-weight model on a laptop isn't. Data stays at the ghat. Photos, questions and family locations never leave the laptop and phones. It's free. No API keys, no subscription — a laptop the family already has. Local languages. Bhojpuri and Maithili rarely get first-class support; open models and open notes let the community improve that. Anyone can swap or tune the model — one line in config/settings.toml — or rewrite the notes for their own family's customs. Bhojpuri and Maithili are beta. The model may slip into Hindi; Hindi is the default. Speed: on a laptop CPU, reading a photo can take tens of seconds, and phones take turns. GPS near water and crowds is often off by 5–20 m — so the landmark photo and note are always on screen. Not a lost-person tracker, and not an authority. Official announcements, police, volunteers and elders come first. Not field-tested yet with real Gemma output — the automated tests use a stand-in model and emulated GPS/compass; the outdoor test is next. Best Use of Gemma — Gemma 3 4B reads the notices and writes the explanations and answers, fully offline, on one laptop. Code (MIT): https://github.com/ThePlator/Chhath_Ghat_Guide- — setup, the design docs (HLD/LLD), and the field-test sheet are in the repo. Checked ritual notes, Bhojpuri/Maithili corrections and sample notice photos (without identifiable people) are especially welcome. छठी मइया सबके मनोकामना पूरा करस। 🌅 — Sameer