Smart India Hackathon 2026 · Problem Statement 26102

Ministry of Statistics and Programme Implementation · Data Informatics & Innovation Division

Software · Smart Automation · Team Technoverse

AI flags. A human decides. Every signal is a risk-prioritised prompt for review — never a finding of fraud.

---

The problem

MPLADS moves ₹5 crore per Member of Parliament per year through thousands of works, across every district in the country, executed by hundreds of implementing agencies. The eSAKSHI portal records all of it faithfully — recommendations, sanctions, staged vendor payments, asset photographs, completion marking.

Recording is not the same as reading. Nobody can read it all, so the hard questions go unasked:

🕐 Which works are stuck?	A work must be completed within one year of sanction. Nobody is watching the clock on thousands of them at once.
💸 Is the money ahead of the work?	Payments are released in stages. A work 20% built and 85% paid looks exactly like one that is fine — until someone compares the two numbers.
📷 Was anything actually built?	Agencies upload asset photographs at each payment stage. A stage released with nothing on record is invisible unless somebody opens that work.
👻 Finished, but invisible	Only works an agency has marked complete show as completed. The portal itself notes that districts are "continuously pursued" to get this done — completed assets sit unrecorded for months.
🔁 Paid for twice?	The same asset, same locality, recommended again months later under slightly different wording.
📅 The March rush	A year's unspent balance committed in the final fortnight, at a rate no ordinary year would produce.

And the one no rule can catch: a work costly but just under the outlier threshold, slow but just inside the year, paid ahead but just under the gap, on a payment schedule no rule examines. Nothing fires. Everything comfortably under every limit — which is what deliberate gaming actually looks like.

---

The solution

Three layers of evidence, in order of how much each one needs — and a human at the end of all of them.

Layer	What it does	Needs
8 rule detectors	Every one traces to a real step of the eSAKSHI workflow and cites the rule it applies. Overdue completion · payment ahead of progress · cost overrun · entitlement breach · missing asset evidence · complete-but-unmarked · duplicate works · year-end clustering	Nothing
2 statistical tests	Cost outliers by median-absolute-deviation against comparable works; agency concentration by binomial tail probability — is this split surprising?, not is it above 45%?	Nothing
IsolationForest	11 engineered features. Finds works unusual in combination — precisely the ones a rule-by-rule check cannot see	Python service
Delay-risk model	Will this running work miss the one-year mark? Uses only what was known at sanction, so it cannot read the outcome it predicts	Python service

The model service is optional by design. ML_MODE=rules skips it; ML_MODE=ml calls it and treats any failure as "no model signals this run". 134 of 174 alerts need no Python at all. An oversight platform that goes dark because a model server restarted is worse than one with no model.

---

Why we are different

🧾 Every alert is a reviewable case Not a score. Not a red flag. Four things, always: The reason, in a sentence, with real figures The rule, stated in words, with its threshold The records it read, breaching row marked The score, itemised — each weighted component, its basis, its contribution, adding up in front of the reader An officer should be able to disagree from that page alone.

🔒 RBAC is a query filter, not a hidden button scopeFor(user) returns a where fragment carried by every query. A missing jurisdiction yields DENY_ALL — the failure mode is see nothing, never see everything. Proved by forging the hidden alertId in a live browser to reach another district's alert. The server refused; the foreign alert stayed untouched. That test ships.

🧭 We publish our own mistakes The delay model first scored AUC 0.98 — a leak, not a result: days past the deadline was in its input. It was reading the answer. Rebuilt on sanction-time attributes only, the honest figure is 0.75, and a test now blocks the leak returning. An accessibility pass caught severity badges at 2.70:1 contrast. A duplicate detector matched boilerplate at 91%. All fixed, all written down.

⚖️ Nothing accuses anyone No alert closes itself. Every state change carries a name. Three of four actions require a written reason. There is no transition back to "untouched" — a decision is answered by recording another, never by erasing it. npm run tune reads reviewer outcomes back, and changes nothing. A detector that retunes itself from reviewer behaviour learns to stop reporting whatever is inconvenient.

---

Architecture

flowchart LR
    subgraph SRC["eSAKSHI workflow"]
        direction TB
        E["Entitlement<br/>₹5 Cr / MP / year"] --> R["MP recommends<br/>& earmarks"]
        R --> S["District sanctions<br/>+1 year deadline"]
        S --> X["Agency executes<br/>staged payments"]
        X --> V["Asset photographs<br/>per stage"]
        V --> M["Agency marks<br/>complete"]
    end

    SRC --> DB[("PostgreSQL<br/>Prisma")]
    DB --> F["Feature engineering<br/>TypeScript · defined once"]
    F --> RL["8 rule detectors"]
    F --> ST["2 statistical tests"]
    F -.optional.-> ML["FastAPI · scikit-learn<br/>IsolationForest + delay risk"]

    RL --> Q["Scored, explainable<br/>alert queue"]
    ST --> Q
    ML -.-> Q
    Q --> H{{"Human officer<br/>reviews"}}
    H --> A["AlertAction + AuditLog<br/>append-only"]
    A --> FB["DetectorOutcome<br/>feedback loop"]
    FB -.a person reads it.-> RL

    style ML stroke-dasharray: 6 4
    style H stroke-width:3px
    style Q stroke-width:2px

Loading

Feature engineering stays in TypeScript, shared with the rule detectors, so there is one definition of "payment-to-progress gap" in the codebase. The Python service receives work ids and numbers — no scheme logic, no names, no personal data.

---

One filter, five jurisdictions

All five roles call the same scope-filtered queries. Only the Scope differs — there is no "national mode" switch to get wrong.

Role	Dashboard answers	Can act on alerts
Ministry / MoSPI	Which states carry the most risk?	✅
State Nodal Authority	Which district do I chase?	✅
District Authority	What can I fix today?	✅
Member of Parliament	What did my entitlement buy?	❌
Implementing Agency	What do I still owe?	❌

An agency marking its own missing-evidence alert as "explained" would make the entire audit trail worthless — so it cannot.

---

Results

	Rule detectors	Statistical	IsolationForest
Precision	100%	100%	not reported — see below
Recall	100%	86% / 100%	100% on multivariate-only cases
Alerts	118	16	40

968 works · 2,162 payments · 4,274 evidence files · 146 ground-truth labels · 6 states · 36 districts

What that 100% is worth. The seed and the detectors were built to the same definition of each rule, so a perfect score is what two correct implementations look like. It is a regression test, not an accuracy claim — and the README says so rather than leading with the number.

npm run eval deliberately reports no precision for the IsolationForest. It is pointed at cases no rule covers, so scoring it for failing to reproduce the rule labels would report 0% for doing its job. What is reported instead: of 12 works planted as unusual-only-in-combination, it found all 12, and 7 of its 8 solo flags were exactly those.

Suite	Tests	Covers
npm test	79	RBAC isolation · detectors · dashboards · review state machine · ML leakage guards
npm run test:e2e	21	Review workflow through a real server action · export scoping · WCAG 2.1 AA on every page · keyboard · हिन्दी
npm run ml:test	12	Model service — refusing to guess on thin data, explanations matching what drove each score

---

Data provenance — read this first

⚠️ The dataset in this repository is synthetic.

It is generated locally by prisma/seed.ts and labelled as synthetic in the database, in a banner on every authenticated page, and here.

Real	State, district and constituency names — so the platform looks and behaves like the deployed thing
Fictional	Every Member of Parliament, official, implementing agency and vendor. Every rupee figure, work, payment and date
Not reproduced	Any official MPLADS record. No figure here is sourced from MoSPI

Planted anomalies are attached only to fictional entities. No real person or organisation is associated with any risk signal in this build — attaching invented fraud signals to real names would be defamatory, and no demo is worth that.

Coverage gap. eSAKSHI holds MPLADS data from 1 April 2023 onward. For the 17th Lok Sabha, FY 2019-20 to 2022-23 is not on the portal, and Rajya Sabha details are unavailable before FY 2023-24. The seed mirrors that window rather than fabricating earlier history, and the UI states the gap wherever history is shown.

Swapping in real data. The detectors read Prisma models, not the seed. An eSAKSHI feed plugs in by writing to the same tables and inserting a DataSource row with kind = REAL. Every figure in the UI renders its source and date from that row, so real and synthetic cannot be confused.

---

Run it

Requires PostgreSQL and Node 18+. No network access needed after install — the font is self-hosted and the app reads only from the local database.

npm install
cp .env.example .env            # set DATABASE_URL and SESSION_SECRET
                                # openssl rand -base64 48
createdb satarkai
npm run db:push                 # apply the schema
npm run db:seed                 # generate the synthetic dataset
npm run detect                  # run the detectors — no Python needed
npm run dev                     # http://localhost:3000

Optional: the model layer

npm run ml:setup     # venv on Python 3.13 + pinned dependencies
npm run ml:serve     # in a second terminal
npm run detect:ml    # now includes IsolationForest and delay forecasts

Python 3.13 specifically — 3.14 has broken pydantic builds. npm run ml:setup pins it.

Demonstration accounts — password satark@2026

Email	Role	Sees
ministry@mospi.demo	Ministry / Central Nodal Agency	Every state
sna.gj@demo.gov	State Nodal Authority	Gujarat, all districts
district.gj-ahd@demo.gov	District Authority	Ahmedabad district
mp.gj@demo.gov	Member of Parliament	Own recommended works
ia.gj-ahd@demo.gov	Implementing Agency	Own assigned works

Substitute the state code (gj, mh, up, tn, wb, as) for other states. Each account sees only its own jurisdiction — that is the point of them.

All commands

Command	Does
npm run dev / build / start	Next.js
npm run db:push / db:seed / db:reset	Schema and synthetic data
npm run detect / detect:ml	Run the detectors, reconcile the alert queue
npm run eval	Score the detectors against planted ground truth
npm run check:baseline	Verify the seed's clean baseline and chronology
npm run tune	Read the reviewer feedback loop back
npm test / test:e2e / ml:test	The three test suites
npm run ml:setup / ml:serve	The Python model service
npm run typecheck / lint	tsc --noEmit / next lint

---

Stack

Next.js 14 · TypeScript (strict) · Tailwind · PostgreSQL · Prisma · Recharts FastAPI · scikit-learn · Python 3.13 · Playwright · Vitest · axe-core

Session cookies signed with jose (httpOnly, secure in production, sameSite=lax, configurable expiry); bcrypt password hashing; no secrets in the repository.

Accessibility is checked, not claimed. axe runs against every page in the e2e suite and fails the build on any violation — currently zero across all seven pages plus the work drill-down. It caught the severity badges failing contrast as small text on its first run; no amount of looking at the page would have.

Charts follow a validated palette. Colour is assigned by the job it does, run through an OKLCH/CVD validator rather than chosen by eye. Severity colour stays reserved for risk — a chart never borrows it for "the third line".

---

Honest limitations

We would rather you read these from us than find them yourselves.

• The dataset is synthetic. Detector accuracy on it says something about the detectors and nothing about real MPLADS execution.
• The 100% rule-detector figure is a regression test, not a real-world accuracy claim.
• The delay model's AUC (0.75) describes how well it recovers relationships this project wrote into the seed — agency reliability, work-type complexity, monsoon slippage. Real delay has its own structure.
• The delay model sees nothing after sanction, because this dataset records only a work's current progress. A real eSAKSHI feed carries staged progress updates.
• Thresholds other than the 365-day rule are review triggers chosen for this build. They need tuning against real reviewed outcomes.
• IA_CONCENTRATION and COST_OUTLIER are statistical signals. A dominant agency may simply be the only one in the district competent to do the work.
• Hindi covers the interface, not generated content. Alert reasons stay in English until the detectors emit message keys rather than finished sentences. The Hindi has not been reviewed by a native administrative-Hindi speaker.
• NOTIFY_MODE=console is the default and no SMTP transport is implemented, so email is recorded as skipped, not sent. SMS and IVR are deliberate stubs that record intent.
• Anomalies are spread evenly across all 36 districts, so a single district dashboard shows only a handful. Realistic, but it makes the District role look emptier than a real one would.
• No detector output is evidence of wrongdoing.

---

Further reading

Document	What it covers
docs/ENGINEERING.md	The full engineering account — every design decision and the mistakes behind it
docs/SCHEME.md	The MPLADS workflow, and the citation behind every rule
ml/README.md	The model service: why unsupervised, why ablation, why so few features

Team Technoverse · Smart India Hackathon 2026 Problem Statement 26102 · MoSPI · Data Informatics & Innovation Division

_{Reference portal — mplads.mospi.gov.in}