Dashboard
Route Optimised inspection Map

M.A.R.G.A: MPLADS Analytics for Risk-Aware Governance & Assistance

Risk Detected, Trust Protected.

Built by Team Code of Chaos

What inspired us

Public money meant for schools, water plants and roads sometimes never becomes a school, a water plant or a road. It becomes a line in a ledger.

One CAG audit in Andhra Pradesh showed this clearly. 3,913 works were reported complete, but only 2,843 were physically verified. In another state, the reporting gap was ₹29.5 Cr (₹54.41 Cr reported vs ₹24.90 Cr in project ledgers).

The usual response is an audit that arrives years later, after the money is gone. We asked a different question: what if the system could flag the problem before the sanction, or while the work is still happening?

That became M.A.R.G.A, a low-power, AI-driven risk detection and decision-support platform for the MPLAD Scheme.

What it does

M.A.R.G.A follows a work through all five stages of its life, and each stage has a tool built for its problem:

Stage	Problem	M.A.R.G.A solution
1. Pre-Sanction	Inflated or prohibited proposals	Smart Proposal Checker (NLP rules + AI cost scoring)
2. Execution	Fake photos, unverified progress	Marga Eyes, a field app with 3-tap geo-tagged updates
3. Completion	Funds out of step with work	Fund & Milestone Checker (reconciliation, auto-generated UCs)
4. Inspection	Scattered, hard-to-plan visits	Smart Inspection Planner (risk-ranked, optimised routes)
5. Oversight	No unified view	Dashboard + Ask MARGA, role-based views with an AI chatbot

Eight features power these stages:

Cost & Time Predictor: predicts a fair cost and timeline from base rates, materials and location.
Inflation Catcher: flags offices whose projects are consistently overpriced against the state average.
Banned Work Scanner: catches disguised prohibited projects (temples, private property, personal names).
Ghost Work Detector: compares site photos against reported completion to catch fake progress and empty sites.
Fake GPS Blocker: verifies GPS and blocks fake photos within a 15 m radius.
Dual-Lock Payment Guard: freezes payouts if payment outpaces work by 15%, pending approval.
Citizen QR Code Audit: a QR plaque shows the contractor and warranty and lets citizens report issues.
Inspection Router: links nearby high-risk sites so more projects are visited in fewer trips.

How we built it

Pipeline: Ingest (CSV, XLSX, PDF, photos) → Clean and normalise → Validate and risk-score → Reconcile records → Route inspections → Dashboards and alerts → Stakeholder action → Feedback into the model.

Tech stack

PyTorch for deep learning models that verify photos and work stages
scikit-learn for anomaly detection, plus BallTree nearest-site search for short inspection routes
LightGBM to benchmark costs and catch inflated billing
SHAP (TreeSHAP) so every cost flag comes with a feature-by-feature explanation
Bhashini for alerts in the officer's own language, including voice
WhatsApp to deliver alerts inside officers' daily workflow
React for role-based dashboards for MPs, District Authorities, Implementing Agencies, State Nodal Authorities, MoSPI and CAG

The platform is designed to plug into existing government infrastructure: PFMS/SNA 2.0, ISRO NavIC, Aadhaar e-Sign, DigiLocker/NAD, PM GatiShakti and CPGRAMS.

The maths behind the alerts

To avoid flooding officers with noise, an agency's new work is flagged only when its cost is a statistical outlier against that agency's own last 50 works:

$$Z = \frac{x - \mu_{50}}{\sigma_{50}}, \qquad \text{flag if } Z > 3$$

The payment guard is a simple rule:

$$\text{Paid}\% - \text{Work Done}\% > 15\% \;\Rightarrow\; \text{freeze payout pending approval}$$

Why it is viable

Lightweight: tree models score a work in under 10 ms on an ordinary CPU, with no GPU needed for the cost and NLP checks.
Live from day one: rules and NLP check 15+ Annexure II clauses with zero training records, and ML joins in as data grows.
Field-ready in 3 taps: photo, status and amount spent. That is about 100 photos per district per month at roughly 300 KB each.
Near-zero cost: about 270 GB of compressed geo-photos at ₹1.9/GB is roughly

$$270 \times 1.9 \approx ₹513 \text{ per month nationwide.}$$

No rule changes: it automates mandates that already exist: the 10% District and 1% State inspections, UCs and the 90-day sanction window.

Impact

$$\text{Verification gap closed} = \frac{3913 - 2843}{3913} \approx 27.3\%$$

$$\text{Inspection trips saved} = \frac{40 - 15}{40} = 62.5\%$$

Verifying a random 10% quota across a 13,000 sq km district takes 40 scattered trips. Anchor-based spatial clustering reaches the same quota in 15.
Fewer trips mean less fuel and fewer emissions, and paperless UCs and digital asset registers reduce paperwork.
Each stakeholder sees exactly what their role needs, backed by a tamper-proof evidence trail.

Challenges we faced

False alarms leading to ignored alerts. If everything is flagged, nothing gets read. We settled on the $Z > 3$ rule over an agency's last 50 works so that only genuine outliers reach an officer.
Evidence tampering. A fraud-detection tool is only trusted if its own records can't be quietly edited. Every flag, photo and officer action is SHA-256 hash-chained, so editing or deleting any past record breaks the chain and is detected.
Fake photos and weak networks. Field officers work with patchy connectivity. We combined geo and time stamps, duplicate-photo checks and offline sync, and added a rule that 90 idle days triggers a show-cause notice.
Policy approval and adoption. Hard locks on payments are a big step for any department. We designed an advisory mode first, reading only from the eSAKSHI feed, with hard locks enabled only after MoSPI/State approval.
Making it usable for non-technical officers. A good model is useless if officers can't use it. This pushed us toward simple dashboards, a three-tap geotagging app, and alerts in the officer's own language through familiar tools like WhatsApp.

What we learned

The hardest part of fraud detection is trust, not accuracy. Explainability (SHAP), tamper-proof logs and a gentle advisory-first rollout matter as much as the model.
Simple and lightweight wins. Rules plus tree models on a CPU were enough to be useful from day one, without waiting for big training datasets.
Smart routing can reduce the cost of oversight as much as better detection can.
Building around existing mandates (10% and 1% inspections, UCs, the sanction window) makes adoption far easier than asking for new rules.