PRAHARIMPLADS Integrity Intelligence

Ministry of Statistics & Programme Implementation · Members of Parliament Local Area Development Scheme

Every rupee leaves
a pattern.

PRAHARI reads the entire MPLADS record — sanctions, releases, bills, work progress and asset evidence — and learns what normal looks like for each category, district and agency. Then it surfaces what does not fit: cost overruns, duplicate sanctions, payments without progress, threshold-splitting, idle funds and vendor capture, each with the evidence that produced the score.

Open the command centre Inspect a live alert queue

Every figure, name, agency and vendor on this prototype is procedurally generated for demonstration. No real MPLADS record or Member of Parliament is depicted, and nothing here is an assessment of any actual person or body.

ALT-26-0010 · Mysuru, KA · Near-duplicate work recommendationALT-26-0011 · Mysuru, KA · Work outside permissible categoryALT-26-0026 · Muzaffarpur, BR · Cost overrun against sanctioned estimateALT-26-0027 · Muzaffarpur, BR · Near-duplicate work recommendationALT-26-0028 · Muzaffarpur, BR · Sanction splitting below approval thresholdALT-26-0032 · Ganjam, OD · Asset geo-tag outside sanctioned wardALT-26-0097 · Rajkot, GJ · Sanction splitting below approval thresholdALT-26-0102 · Nashik, MH · Payment without physical progressALT-26-0109 · Bhavnagar, GJ · Vendor concentration in a constituencyALT-26-0132 · Kalaburagi, KA · Cost overrun against sanctioned estimateALT-26-0134 · Kalaburagi, KA · Asset geo-tag outside sanctioned wardALT-26-0057 · Coimbatore, TN · Unit rate above district schedule of ratesALT-26-0009 · Mysuru, KA · Payment without physical progressALT-26-0036 · Alwar, RJ · Near-duplicate work recommendationALT-26-0010 · Mysuru, KA · Near-duplicate work recommendationALT-26-0011 · Mysuru, KA · Work outside permissible categoryALT-26-0026 · Muzaffarpur, BR · Cost overrun against sanctioned estimateALT-26-0027 · Muzaffarpur, BR · Near-duplicate work recommendationALT-26-0028 · Muzaffarpur, BR · Sanction splitting below approval thresholdALT-26-0032 · Ganjam, OD · Asset geo-tag outside sanctioned wardALT-26-0097 · Rajkot, GJ · Sanction splitting below approval thresholdALT-26-0102 · Nashik, MH · Payment without physical progressALT-26-0109 · Bhavnagar, GJ · Vendor concentration in a constituencyALT-26-0132 · Kalaburagi, KA · Cost overrun against sanctioned estimateALT-26-0134 · Kalaburagi, KA · Asset geo-tag outside sanctioned wardALT-26-0057 · Coimbatore, TN · Unit rate above district schedule of ratesALT-26-0009 · Mysuru, KA · Payment without physical progressALT-26-0036 · Alwar, RJ · Near-duplicate work recommendation

Works under continuous watch

168

sanctions in the demo corpus

Value monitored

₹84

crore sanctioned

Open anomalies

128

46 at critical tier

Districts covered

124 implementing agencies

01 — The scale

Thousands of small works, spread across every district.

MPLADS moves money in many small sanctions rather than a few large contracts. That shape defeats manual scrutiny: no reviewer can hold the comparison set for a ₹14 lakh anganwadi in one block against the four hundred comparable works that would tell them whether the rate is reasonable.

02 — The blind spot

Irregularity hides in the ordinary, not the exceptional.

Fund misuse rarely looks dramatic. It looks like a sanction just under a delegation limit, a bill cluster in the last fortnight of March, a completion photograph from the wrong ward, the same asset sanctioned twice under different descriptions.

03 — The answer

Learn the norm, then rank the departures.

Every record is scored against its own peer cohort — same category, district, quarter, agency size. What remains after that normalisation is worth a human's time, and arrives with its reasoning attached rather than as a black-box verdict.

The detection lattice

Twelve detectors, one calibrated score.

Each detector answers a narrow question well. A calibration layer folds their outputs into a single comparable 0–100 score so a road in Barmer and a school block in Murshidabad can sit in the same queue.

D-01Financial

Cost overrun against sanctioned estimate

Final expenditure diverges from the approved estimate beyond the tolerance learned for that work category, district and year.

Gradient-boosted residuals

D-02Financial

Unit rate above district schedule of rates

Per-unit cost of an asset sits in the upper tail of comparable works in the same district and quarter after size normalisation.

Robust z-score on peer cohort

D-03Procurement

Near-duplicate work recommendation

Two or more sanctions describe the same asset at the same location, across MPs, years or implementing agencies.

Sentence embeddings + geo clustering

D-04Procurement

Sanction splitting below approval threshold

A cluster of sanctions to one agency lands just under a delegation limit within a short window - the classic threshold-avoidance signature.

Threshold-proximity density scan

D-05Execution

Payment without physical progress

Funds released and drawn while the geo-tagged photographic and measurement-book evidence shows no corresponding advance.

Isolation Forest on progress-payment gap

D-06Execution

Predicted completion overrun

A forward estimate of the completion date from progress velocity, monsoon seasonality and agency history flags works that will slip.

Survival model + seasonal decomposition

D-07Procurement

Vendor concentration in a constituency

One firm captures a disproportionate share of awards under a single MP or agency, adjusting for the size of the local contractor pool.

Bipartite graph centrality + HHI

D-08Financial

Fiscal year-end payment burst

Abnormal clustering of releases and bills into the closing weeks of the financial year relative to the work's own execution curve.

Point-process burst detection

D-09Financial

Idle funds beyond permissible period

Money released to the implementing agency but unspent past the scheme's holding window, with no revised work programme on record.

Rule engine + ageing buckets

D-10Financial

First-digit anomaly in bill values

The distribution of leading digits across an agency's invoices departs from the Benford expectation, a marker for manufactured figures.

Benford chi-square test

D-11Execution

Asset geo-tag outside sanctioned ward

The completion photograph's coordinates fall outside the recommended ward or block boundary by more than the GPS error margin.

Spatial join + haversine outliers

D-12Procurement

Work outside permissible category

The recommended work maps to an item barred by the MPLADS guidelines, or to a trust or society outside the eligible list.

Zero-shot classifier on guideline corpus

From record to redress

The path a suspicious payment takes.

Ingest

Sanctions, releases, bills, measurement books, geo-tagged photographs and utilisation certificates arrive from the MPLADS portal, PFMS and state treasuries.

Reconcile

Records are resolved into a single work identity across agencies and financial years, with vendors, agencies and wards linked into one graph.

Model

Twelve detectors run in parallel — outlier ensembles for money, survival models for time, embeddings for duplication, graph measures for procurement.

Score & explain

Detector outputs are calibrated into one 0–100 risk score, with the feature contributions that produced it attached to every alert.

Act

Alerts route to the right desk with a response clock, evidence pack and audit trail. Every disposal feeds back as a label for the next training round.

Four desks, four views of the same truth

Scoped to the person reading it.

Member of Parliament

Ramanagara 2 constituency

Sees only their own entitlement: what has been recommended, what is stuck, and which recommendations are at risk of being spent badly.

Own works and balance
Slippage forecast
Agency responsiveness
Asset completion evidence

District Authority

District: all constituencies

Works the queue. Every open alert lands here first with the evidence attached and a clock running against the response window.

Alert triage queue
Agency and vendor performance
Sanction-to-release latency
Inspection scheduling

State Nodal Authority

State: all districts

Compares districts against one another, spots agencies whose anomaly rate is out of line, and escalates what the district has not closed.

District league table
Repeat-offender agencies
Escalation backlog
Utilisation certificates

Ministry (MoSPI)

National: all states

Watches the scheme as a whole: national risk index, where money is idle, and which anomaly families are growing quarter on quarter.

National risk index
Fund-flow leakage
Detector drift and precision
Policy-level trend analysis

Explainability is not optional here

An alert that cannot justify itself is noise.

A risk score that no one can interrogate will not survive contact with a District Collector, and it should not. Every PRAHARI alert carries the ranked feature contributions that produced it, the peer cohort it was compared against, the underlying records, and a confidence figure. Reviewers can disagree — and their disposal becomes a training label, so precision improves where the scheme actually operates.

SHAP-style attributionPeer cohort disclosedReviewer feedback loopFull audit trail