←BaldanAll worksLet's talk

Case study · GatherInsight · field apps · 2026

Three field apps that keep working without signal.

GatherInsight's web platform plans the research; these phones do the fieldwork. Surveyors interview, count traffic and map places, often far from any signal, and every record comes back with its proof: where, when, and what was said.

Expo · React NativeSQLite offline queueNext.js · PostgreSQL backend
09:41●●●

Input

Offline · 3 entries waiting to send

Customer satisfaction 2026

Ready offline · saved 08:12

12 / 20

Household internet use

Ready offline · saved 08:12

7 / 15

Store visit audit

Ready offline · saved 08:12

18 / 18

Consent

Voice recorded after consent

DeclineAgree
Study & Research
09:41●●●

Traffic count

Site JL-07 · 41 m away

02:14:36 / 12:00:00

Interval 9/48 · 06:32 left

Direction A · 312Direction B · 286
MC412motorcycle
LV156car, pickup
HV21bus, truck
UM9bicycle, cart
UndoPause
Mobility
09:41●●●

Area · Parcel 14

Walking the boundary

20 m

2,418 m²

area

23

corners

±4 m

accuracy

96%

inside zone

PauseFinish
Geo

Phones are drawings of what each app does, made from its code. Real screenshots are coming.

01 · Three apps

One platform, three jobs in the field.

Separate apps, so each install carries only what its job needs. They share one codebase: the same database layer, sync pattern, GPS code and sign-in.

Study & Research

Questionnaires and interviews in the field.

  • Picks the respondent from the sample frame, quotas or a household draw
  • Consent first; voice recorded only after a yes
  • Results and refused entries, with the reason, on the phone

Mobility

Traffic counts at a site, 12 hours by default.

  • One big tile per vehicle class, per direction
  • Every tap stored as an event, in 15-minute bins
  • Pushes the count live to the office every 45 s

Geo

Six ways to map: points, lines, areas, samples, inspections, incidents.

  • Geometry recorded by walking with GPS, cleaned on the phone
  • Plans drawn on the phone, so maps work with no signal
  • The server measures again; the phone never gets the last word

02 · One interview

Gate by gate, and only the gates a survey needs.

Even a visit that goes nowhere (declined, not eligible, nobody home) is recorded, because a research team needs its response rate as much as its answers.

  1. 1

    Pick the respondent

    From the sample frame (main, reserve, declined, unreachable), a crossed quota or a systematic sequence. The GPS check runs before anything starts.

  2. 2

    Household draw

    When the survey picks one member per house: everyone's birthday goes in, the next birthday wins. Only if the survey needs it.

  3. 3

    Screening

    The criteria as a checklist. Not eligible? The visit is still recorded, just not counted.

  4. 4

    Consent

    A no is recorded too. A yes starts the voice recording when the survey asks for one.

  5. 5

    The questions

    One page per section. Required answers and ranges checked before Next; branching rules pick the next page; answers on skipped pages are dropped before sending.

  6. 6

    Review and send

    GPS still on? End position taken, the last audio piece queued, then the interview joins the send queue and the phone says Sent, Refused (with the reason) or Saved.

03 · Offline first

No signal is the normal case, not the error.

Surveys download ahead of time. A finished interview goes into a queue on the phone with an id the phone made itself, and the queue drains oldest first when the signal returns, and every 30 seconds. Lose the signal, finish a few interviews, then make the server misbehave.

Phone · SQLite queue

No signal
  • Interview #1 · CSAT

    id 5e98…2adb

    Waiting to send
  • Interview #2 · NPS

    id 7d87…c3f4

    Waiting to send

2 waiting · ids are made on the phone, so a resend can never create a second row.

Server · responses

0

rows stored, one per interview id

No signal. Two interviews are saved on the phone, waiting.

04 · GPS rules

The office sets the rules, the phone enforces them.

Per survey: capture at the interview or continuously while the app is open, required or not, lock the whole app until GPS is on, what happens outside the assignment area, an accuracy limit, how old a fix may be, whether fake GPS is refused. A surveyor can be paused, and the phone picks up new rules within minutes.

assignment · 300 myou · ±40 mdashed ring = area + margin (40 m)

blocked · OUTSIDE_AREA

You're outside the assignment area. Move closer, then try again.

The office sets these per survey. Tap the three boxes to change them.

The server judges each interview by the rules the phone had when it started, so changing a rule mid-day flags interviews instead of throwing them away. It also compares the phone's clock with its own and the start position with the end position, and flags what doesn't add up.

05 · Voice evidence

What was actually said, in small pieces.

  • Only after consent

    Recording starts on the consent screen's Agree, never before, and only when the survey has voice notes on.

  • Two-minute pieces

    A new file every two minutes, AAC at 16 kHz mono and 32 kbps: about a quarter of a megabyte a minute. A weak connection only ever has a small file to move.

  • Each piece on its own

    Pieces queue and upload separately, numbered in order. Sending the same piece twice is harmless, and the file is deleted from the phone only after the server confirms it.

  • Tied to the interview

    A session key made at consent links every piece to the interview, so the office can listen next to the answers.

one interview, 5 min 12 s

Consent: yes→recording

0000.m4a

uploaded

0001.m4a

uploaded

0002.m4a

waiting

0:002:004:005:12

The last piece is queued at Submit. Leaving the interview or starting another one ends the recording, so audio never runs into the next respondent.

06 · Mobility

A traffic count that can't lose a tap.

Twelve hours at a roadside, one tap per vehicle. Each tap is written to the phone's database before the screen even updates, so a crash or a flat battery loses nothing. Tap a few vehicles, undo one, jump the clock.

06:07:00

started 06:07 · interval 15 min

Event log · saved before the screen updates

Tap a vehicle class.

Undo never deletes. The office sees every correction, and too many is a review flag.

15-minute bins · computed from events

06:00–06:150running · 0:00

Bins sit on the clock, so counts from different sites line up. A partly watched bin says so instead of looking quiet.

A pause needs a reason (rain, rest, road closed) and paused time isn't counted as an empty road. Ending a session runs the same form checks as the server, then the server reviews it: distance to the site, how long it was watched, empty intervals, the share of undos, and intervals far out of line.

07 · Geo

Six ways to map, measured twice.

Maps on the phone are plans drawn with SVG from the downloaded work pack, not map tiles, so they work in a field with no signal. The geometry code is copied from the server, so the phone shows the warnings the server will raise before anything is sent.

walking an area: raw → kept → thinned

every fixkept, metres apartthinned

Thinning uses Douglas–Peucker; badly inaccurate fixes are thrown out on the way. A walk in progress is saved every few corners, so closing the app doesn't lose it.

Points

About ten GPS readings in twenty seconds. Readings far from the median are dropped, the rest averaged, weighted by how accurate each one was. Too close to a similar point? The phone asks.

Lines

Walk or drive along a road or drain. Points closer than a few metres are skipped, the line is thinned with Douglas–Peucker, and its ends can snap to existing segments.

Areas

Walk the boundary, stand at each corner, or draw with a reason. Corners that were walked past quickly are recovered from the buffered fixes. Area and overlap are measured on the phone and again on the server, which has the last word.

Samples

Go to a planned point: arrived, moved a little (give a reason), or too far (can't save). Take repeated readings; noise can be averaged as decibel energy rather than as plain numbers.

Inspections

Revisit features in rounds and mark them confirmed, changed, missing or unreachable. A change is a proposal that a person, or the rules, approve.

Incidents

Report with a severity that sets the response time and an affected radius. Never refused by a quality rule: a late report is better than none.

08 · Under the hood

Small rules that keep a field team's data honest.

Edits during upload survive

In the Geo apps every save bumps a revision number. The server's reply only marks a record clean if it's still the revision that was sent.

One pass at a time

Sync never overlaps itself, so the same entry is never sent twice at once and the order holds.

Photos from the camera

No gallery. A photo that can't upload yet waits inside the answer and goes before the interview does.

Signed in, offline

The token sits in the phone's secure store, so a surveyor who signed in once can work all week without signal.

The office can lock the app

If a survey requires it, the whole app shows only “turn on GPS” until location is on, and an office pause releases it.

Too-old apps are told

Every request carries the app version; the server can require an update.

Two languages

Indonesian and English, switchable on the login screen; server messages are translated back on the phone.

Tested like a phone

Tests run the real queue and sync code on in-memory SQLite against a dev server, simulating no signal, lost replies and hangs.

Expo SDK 57React Native 0.86expo-routerexpo-sqlite (WAL)expo-locationexpo-audioexpo-secure-storeNetInforeact-native-svgNext.jsPrismaPostgreSQLCloud RunCloud Storage