Development of an AI-powered system to detect anomalies, fraud, and inefficiencies in MPLAD Scheme implementation regd.
MoSPI · Miscellaneous · Software
Genuine public administrative data at national scale with verifiable findings is rare here, and audit reports give you real irregularity patterns to encode — just frame it as outlier detection for review rather than fraud detection, and keep identifiable constituencies out of the demo.
What it actually is
Members of Parliament recommend local development works — roads, community halls, water facilities — funded from a scheme running thousands of projects nationwide through many implementing agencies. With that volume nobody can spot the works that cost far too much, never got built, or were sanctioned twice. The ask is a system that finds those patterns in the scheme's own data.
What to build
An anomaly detection layer over the scheme's published works data, framed honestly as outlier detection rather than fraud detection since no confirmed fraud labels exist: cost-per-asset outlier detection comparing similar works within and across districts so a community hall costing several times the district norm surfaces; near-duplicate work detection catching the same work description sanctioned at the same or adjacent locations more than once; execution stall detection finding works sanctioned and funded long ago with no progress recorded; utilisation pattern analysis flagging agencies or districts whose expenditure and completion behaviour departs from peers; and a reviewer interface where every flag is explained in terms of the specific comparison that produced it, with the pattern library grounded in the irregularity types that public audit reports on this scheme have already documented.
Smallest thing that wins the room
Show the ten highest-scoring flagged works with the reason beside each — this asset costs four times the district median, this description appears twice at the same location, this work has been funded and stalled for three years — and let a judge click into one and see the comparison set that produced it.
How crowded this one gets
A guess, projected from the 2025 statements — the last year where both the submission counts and the winners were published.
Quieter than 24% of the 226 · #173 of 226 by expected field
A normal-sized field. Your idea has to be good, not miraculous.
Why: central ministry statements sat below the average.
This is a guess, not a fact
Nobody has published 2026’s numbers yet. This is an analysed estimate from last year’s pattern, so please do not take it as the truth — check the live counter on the SIH portal before you decide anything. The range covers the middle half of likely outcomes, so one statement in two lands outside it. Entry closes at 500 ideas per statement, so no range goes past that — a statement that reaches the cap fills and shuts rather than drawing an unlimited crowd. The model reads only three things a team can see before choosing — software or hardware, the theme, and what kind of body posted it — and those explain about a quarter of the variation in last year’s field sizes (R² 0.25 on held-out statements). Trust the band more than the number, and the ordering more than either. It cannot see how good your idea is, which is the part that actually decides it.
The scores
The number is the shorthand. The line under it is the reason.
Acceptance potential
4/5Real public administrative data at national scale is rare on this portal and it is the whole game here, the findings are concrete and verifiable against published records, and public audit reports on this scheme document the irregularity patterns you can encode — so your detectors can be grounded in documented reality rather than invented.
Feasibility
4/5This is one of the few statements on the portal backed by real public administrative data at scale — the scheme portal publishes works, sanctioned amounts, expenditure and completion status across constituencies and districts, so you are analysing genuine government records rather than a synthetic set.
Innovation scope
4/5The statement names no method, no anomaly definition and no model, so what counts as an irregularity, how you construct comparison groups and how you avoid flagging legitimate variation are all yours to determine and are the substance of the work.
Clarity
3/5The background, description and expected solution sections restate essentially the same paragraph three times, and the statement never defines what an anomaly is or supplies a single confirmed irregular case, so you are setting the detection criteria yourself.
Effort
HeavyAcquiring and cleaning the works data, constructing meaningful comparison groups, building several distinct detectors, scoring and ranking flags, and building an explainable reviewer interface is four workstreams with the comparison-group construction being the analytically hardest.
Demo-ability
MediumA ranked list of flagged works with concrete reasons is genuinely compelling to an administrator and each flag is checkable against the source record, but it is a table rather than a visual and needs the comparison logic explained for the ranking to mean anything.
In its favour
- Green flag: The data is real, public and at national scale, which puts this in a small minority of statements here where you analyse genuine records rather than something you generated
- Green flag: Public audit reports on this scheme document the actual irregularity patterns that have been found historically, so you can encode detectors grounded in what has really gone wrong rather than in what you imagine might
- Green flag: Every flag is verifiable — a judge can look up the flagged work in the public portal and see the numbers for themselves, which is a very strong position
- Green flag: Near-duplicate work detection is both technically interesting and immediately convincing, since the same asset sanctioned twice at the same location is unambiguous once you find it
Against it
- Red flag: There are no confirmed fraud labels, so this is unsupervised outlier detection and calling it fraud detection is a claim you cannot support — a flagged work is a work worth looking at, nothing more, and framing it otherwise is both wrong and unfair to the people named
- Red flag: Legitimate cost variation is enormous — terrain, materials, remoteness and timing all move unit costs — so comparison groups have to be constructed carefully or you will flag every hill district
- Red flag: Naming specific constituencies or members in a demo attaches an allegation to real identifiable people on the basis of a statistical outlier, so anonymise the presentation and let the method be the claim
- Red flag: The description repeats itself three times and defines nothing, so your detection criteria are your own and a panel may measure you against different ones
What you will be writing
- cost-per-unit outlier detection within peer groups
- near-duplicate work description matching
- isolation forest on utilisation patterns
- stalled execution detection from progress timelines
- explainable flag attribution to comparison set
- public MPLADS works data ingestion
- Public expenditure oversight
- Anomaly detection
- Development scheme monitoring
Prior art to read before you start
administrative data anomaly detection · duplicate work and cost outlier identification · explainable risk flagging for auditors
Analysed by Claude Opus. Every score above is a judgment call with its reasoning attached — kindly cross-check this against the official statement on the SIH portal before your team commits to it.