Problem Statement
Most sari-sari stores run on memory and a stained notebook. Utang lists get lost, owners don’t know if they earned or lost in a week, and restock is guesswork. Existing POS apps need accounts, OTPs, and always-on data — plus owners fear their sales data goes to a cloud they don’t control.
Proposed Solution
TindahanLedger works like talking to a helper: voice or quick-tap logging of benta, utang, and gastos. On-device NER pulls item, amount, and person from Taglish speech. Stock counts decrement automatically; low-stock and top-utang alerts appear on one weekly page. Everything stays in local SQLite. Optional thermal-printer receipt for customers. Works with zero SIM.
Target Market
- Primary users: ~1.3M sari-sari/market stall owners, especially nanays managing utang.
- Deployers: Cooperatives and youth volunteers who set up phones during market day.
How AI Is Used
- Offline speech NER: Extracts (item, amount, person, utang-vs-cash) from messy Taglish utterances.
- Smart weekly summary: On-device LM writes “Kumita ka ₱2,340; pinakamalakas Ariel; singilin si X” from ledger rows.
- Why AI: Keyword lists break on real speech (“yung malaki, yung kay Nena, lista mo muna”).
Independent Deployment Angle
- No telco API, no cloud, no venue, no headset. Any Android 9+ phone works.
- Data stays with the owner — the privacy pitch that unlocks adoption.
- One-hour training by a student volunteer is enough.
Demo Scenario (3 minutes for judges)
- Airplane mode on. Say: “Benta ₱150 Ariel, utang ni Ate Nena.”
- Ledger row appears with item, amount, utang tag. Add two more by voice.
- Tap “Lingguhang Kwenta” — one page shows kita, top items, sino sisingilin.
- Show low-stock alert for Milo after three sales.
- Close: “Walang account. Walang cloud. Nasa iyo ang datos.”
30-Hour Build Plan
| Hours | Task | Role |
|---|---|---|
| 0–3 | Collect 100 sample utterances, define ledger schema | All |
| 3–10 | Vosk STT + NER extraction prompts, Taglish few-shots | AI lead |
| 3–10 | SQLite ledger, utang aging, stock decrement | Backend |
| 10–20 | Voice-first UI, weekly summary page | Frontend |
| 20–26 | Test with real tindera phrasing, tune thresholds | All |
| 26–30 | Pitch, disclosure | PM / Pitch |
Suggested stack: Flutter, Vosk Tagalog STT offline, on-device NER TinyLM, SQLite. Public data / APIs: None; team-collected synthetic utterances.
Risks & Privacy Notes
- Money errors: Every AI parse shows an editable confirm chip; weekly totals recompute from rows.
- Privacy: Fully on-device; no personal or sales data leaves the phone. Demo uses fictional names only.
- IP: All speech samples synthetic and team-owned; OSS licenses credited.
AI Use Disclosure (draft)
Uses offline STT plus a small on-device model for entity extraction and weekly summaries. AI coding assistants helped with UI boilerplate. All parses in the demo were human-verified against the ledger.
Rubric Scoring
| Criterion | Weight | Score (1–5) | Why |
|---|---|---|---|
| Innovation & creativity | 25% | 4 | Voice utang ledger on-device is new for sari-sari |
| Problem–solution fit | 25% | 5 | Daily pain, zero-trust-of-cloud solved |
| AI application & value | 20% | 4 | Speech NER essential for hands-busy owners |
| Feasibility | 15% | 5 | Very buildable in 30h |
| Impact & scalability | 15% | 4 | Millions of stores; volunteer-led scale |
| Weighted total | 4.30 |