Turning footpaths into schools

1,600 hours a month
went into paperwork
They should have gone into teaching.

Footpathshala runs across zones, centres, volunteers and hundreds of children — on spreadsheets that do not talk to each other. This replaces them with one system that records what happens, then answers questions about it in plain language.

1,600+: hours a month, previously on data entry
1 photo: marks a whole class register
0 images: of children are ever stored

Footpathshala · Dashboard

Hello, Meera 👋

412

children

71%

attendance

volunteers

ATTENDANCE TREND

Numeracy — Building blocks

Delivered 👏 · 45 min

Life Skills — Practice

Coming up ⏳ · Thu

Recognised

18 / 21

faces matched 🎯

Certificate

100 hours 🏅

What it does

Every module writes to one database. That is the whole point — it is what makes the analytics, and UpayGPT, possible at all.

Attendance from one photo

Faces are matched in the browser against the centre roster. The photograph never leaves the device and is never stored — only a 128-number embedding is kept.

Ask in plain language

UpayGPT writes the SQL, runs it on live data, and shows you the query it used. An answer nobody can audit is not an answer.

Geo-verified check-in

Distance from the centre is recomputed on the server, so service hours reflect where a volunteer actually stood.

Curriculum that travels

Build a lesson once, schedule it across centres, and see which lessons were genuinely delivered.

Analytics that name names

Not totals. Which centre is slipping, which children are falling behind, which centres are short of volunteers.

Certificates, issued automatically

Hours accumulate and eligible volunteers receive a serial-numbered certificate without anyone chasing a spreadsheet.

Bring the old records with you

Years of spreadsheets import through a column-mapping step, so the new system starts with the history rather than pretending it does not exist.

Hindi where it matters

Navigation and the controls a volunteer touches in the field are translated — not a machine-translated shell that claims more than it delivers.

UpayGPT

Ask the database a question the way you would ask a colleague.

Nobody at a field NGO should need SQL to find out which centre is struggling. Type the question; UpayGPT writes the query, runs it against live data, and shows you both the answer and the query.

That last part is the design. The generated SQL is always visible and always logged, and it runs under your own permissions — asking a question differently does not widen what you are allowed to see.

Which centres had the lowest attendance last month?

Three centres fell below 55%. Sitabuldi Signal dropped from 78% to 51% over six weeks — the steepest decline in the programme.

12 rows1.8s▾ see the SQL I ran

select center_name,
       round(avg(attendance_rate), 1) as rate
from v_attendance_daily
where session_date >= date_trunc('month',
        current_date - interval '1 month')
group by center_name
order by rate asc
limit 12

These are children

Face recognition, without a database of children’s faces.

Recognition runs entirely in the volunteer’s browser. The class photograph is decoded, measured, and discarded on the device — it is never uploaded, and there is no column in the schema to store it in.

What is saved is a list of 128 numbers per child, which cannot be turned back into a photograph. Every match is shown to the volunteer with a confidence reading, and they can override any row before the register is saved.

1
Photo captured on the volunteer's phone
stays on the device
2
Faces detected and measured in the browser
no upload, no server cost
3
Matched against that centre's roster only
128 numbers, never an image
4
Volunteer reviews and corrects
the model never has the last word

1,600 hours a monthwent into paperwork🗂️They should have gone into teaching.

What it does 🧩