Six tiles, no answer

Three people open this page with a question. None of them can answer it.

MSME Samadhaan is the portal where a micro or small enterprise files a complaint when a buyer pays late — and where the Ministry, 37 state Facilitation Councils and Parliament go to see how the system is doing. An MSME owner wants to know: if I file in my state, against this kind of buyer, will anything happen, and how long will it take? A council chair wants to know which of the thousands of pending applications to hear first. A minister is asked in Parliament how much money is stuck and whether it is getting better. The front page answers none of these. It answers "how many rows are in each table."

Eleven report pages, all tables, no join. So the first job was the boring one: read every table and make the totals agree.

Behind the tiles sit eleven separate reports — by state, by council, by buyer category, by railway zone, by ministry, by ageing — each a long HTML table with two numbers crammed into every cell (a count and a rupee amount, separated by a space). Nothing links them. Read together, they disagree in ways the front page never mentions:

# every cell on the portal holds two numbers as one string: "53322 10339.56" = count, ₹ crore
def split_cell(s):
    parts = s.split()
    return int(parts[0]), float(parts[1])             # count, amount_cr

states = [parse_row(r) for r in statewise_age_report.rows]
assert abs(sum(s.filed for s in states) - welcome_tile.filed) < 100, \
    "state report and front page disagree — publish the gap, don't hide it"   # they do: 2,107 apart

A faithful recreation of the front page as it was in May 2025 — with the ten things wrong with it.

A dashboard is a list of questions with a chart under each. So — the questions, in the order the money cares about.

Money first. Age second. Place third. Then the people who owe it, the councils that clear it, and one honest note about the headline.

Every change above is a rule, not a taste. Here they are, so the next dashboard gets them for free.

The rebuild is worthless if it is a screenshot. The pipeline that keeps it honest is forty lines.

# nightly: read the eleven reports, reconcile, publish (simplified)
REPORTS = ["StateWise_AgeRpt", "CategoryAmtAge_Rpt", "Council_Report", "ReportAllAmount", ...]
frames  = {name: parse_two_number_cells(fetch(name)) for name in REPORTS}

# three reconciliations that have to hold — or the gap gets published as a number, not swallowed
gap_filed  = frames["StateWise_AgeRpt"].filed.sum() - welcome.filed
gap_amount = welcome.claimed_cr - sum(stage.amount_cr for stage in funnel_stages)
gap_cat    = frames["CategoryAmtAge_Rpt"].filed.sum() - frames["ReportAllAmount"].filed.sum()

snapshot = dict(as_of=today, funnel=funnel, ageing=ageing_buckets, states=top_states,
                categories=cats, councils=council_rates, gaps=dict(filed=gap_filed, amount=gap_amount))
publish(snapshot)                                          # the dashboard reads one JSON, never a table
alert_if(pending_over_a_year.amount_cr > last.pending_over_a_year.amount_cr * 1.05)   # the WATCH

Two design choices in that pipeline matter more than the code. Gaps are published, not swallowed — the front page would carry a small line reading "reports disagree by 2,107 applications" instead of pretending they agree. And the alert is on money over a year old, because that is the number a delayed-payment system exists to shrink; everything else is activity.