This project has the richest conversation record, so it doubles as the real example Build Log for 'How I Build.' All quotes below are public excerpts (1–2 key sentences).
← All work
Case 06 / 09
Money Navigation
An app that helps travelers lose less money on every overseas payment — finished in a 48-hour hackathon by directing multiple AI sessions
ReleasedIdea→ship48 hours
2026.10 (prep Oct 7–10 · contest Oct 15–16) / Solo (multiple AI sessions as the team) / iOS · Android
- Problem
- Travelers lose a little money on every payment abroad — ATM fees, exchange, haggling.
- My role
- Planning · multi-session AI direction (PM model) · release
- Stack
- Flutter · Dart · Supabase Edge Functions (Deno) · ML Kit (on-device OCR) · Google Vision/Places/TTS/Translation
Results
- Contest start → GitHub·APK·TestFlight submission
- 48 hours
- Contest start → GitHub·APK·TestFlight submission
- Parallel AI sessions run under a PM model
- 8+ sessions
- Parallel AI sessions run under a PM model
- Public GitHub · APK v1.0.0-7 · TestFlight build 8
- Public release
- Public GitHub · APK v1.0.0-7 · TestFlight build 8
⚡ Idea→ship
48 hours
2026-10-15
First prompt · First working build
+1 day
2026-10-16
First release
+2 days
2026-10-18
Latest
Contest build start → GitHub·APK·TestFlight submission ~2 days (48h). Idea selection (10-03) and prep (10-07) were earlier; security fix redeployed 10-18 (4 active days in total, incl. prep and the fix)
01
02
Idea / Problem
Topic-selection instruction: "List ideas that are unique, commercially viable, and likely to score high with judges — research cases first, and dig deeper to confirm they really don't exist." Original (KO): "유니크하고, 상업성 있고, 심사 점수 높을 아이디어를 사례부터 찾아 나열해줘 — 정말 없는지 더 깊게 파봐."
A map tells you where to go, but nobody tells a traveler how to pay abroad with the least loss. A little money leaks away each time — ATM fees, exchange, haggling, over-charged receipts. I built a travel-payment assistant to cut that, at U-Hackathon 2026 (AI track) in Bangkok.
03
The Plan & Bar I Set
- Demo order = build order: ATM scan → voice haggling → receipt verification → home/assistant
- Quality bar: "Divide zones and give depth like a big-company app (Toss, Wise, Apple Wallet). It must not look AI-made."
- For the 'AI track,' I separated on-device (internal) and cloud (external) AI roles so you can see "this runs on AI"
- Pinned rules in a rules document: "No faking / test at three points / don't decide from what you alone know"
04
Directing AI
Fig. money-nav-01
Running many AI sessions like a team — the PM setup
MePlan & decideTalk only to the PM
PMPM sessionOrders the work · relays reports in Korean
S1
S2
S3
S4
S5
S6
S7
S8
Sub-sessions ×8 (parallel)
Fig. money-nav-05
Directing AI
01 First instruction (prep)
"Read everything first and summarize what you understood. Don't build anything yet — I'll fix what's wrong."
02 Kickoff
"Decide today's build order and show me the plan in 5 lines before starting." — fix scope and rhythm first.
03 Division of labor
while AI wrote code, I pulled the screen designs in parallel.
04 Multi-session
"If there are several sessions, make one the PM to manage order, and report status to me right away." → After that I talked only to the PM session, which relayed every session's report in Korean.
05
Where It Got Stuck & How I Resolved It
- "This isn't hardcoded, right?" (scan verification): The ATM scan looked like it worked on the phone, but it could be faked to the demo numbers, so I had it re-run on photos with different numbers. The real text-recognition output had mixed-up columns; I reinforced the alignment logic and confirmed zero false positives across 51 real receipts.
- Voice-recognition timing: I relayed the exact symptom ("it stops listening before I even speak"), then extended the listening window and filtered out late results, re-testing with quiet/noisy recordings.
- Paid integration blocked: Blocked by settings outside my permission, I said "don't waste time changing policy" and pivoted to what works (free/on-device), holding quality within the deadline.
- Post-launch hardening: After public feedback, I reinforced caller verification on the server functions and redeployed.
06
Results
- Finished the hackathon (no award) — ATM scan, voice haggling, receipt verification, and a travel assistant all working within 48 hours
- Public GitHub repo + APK release (v1.0.0-7) + TestFlight build 8 + demo video
- Zero false positives across 51 real receipt photos
Certificate of Participation — U-Hackathon 2026, AI Track (u-topia), 16 Oct 2026. A participation certificate, not an award.
07
Retrospective
This project showed most clearly that "building solo with AI" is really about running several AIs like a team. Next time I'd set up the PM model and file-ownership rules from the start to reduce session conflicts further.
More screens · 2
Design boards made first in Google Stitch, then the same flows built in the Flutter app.
Next project →
Adonis AI
A camera-based AI personal trainer (hackathon prep) — pivoted to Money Navigation at the event
