How AI Automates PM Kusum Yojana Solar Installation Verification and Fraud Detection

Standard pixel anomaly models compute a modification probability across the full image frame and report any region where the pixel distribution deviates from the expected noise floor. Applied to field photographs, this produces false positives on every GPS banner and timestamp overlay -- which appear on virtually every authentic submission. The exclusion mask pipeline solves this by building a spatial map of benign overlay regions before the forensic pass runs, using edge-density heuristics to detect rectangular metadata banners and text bounding box positions from the OCR stage to identify watermark zones. The anomaly analyser evaluates only the pixel regions that remain after subtracting this mask. The result is high sensitivity to structural fraud -- cloned panel regions, inpainted personnel -- precisely where it matters, with no false positives on camera overlays. This component is the one a naive implementation always skips and the reason naive implementations cannot be deployed in production.

Field photographs capture text that mixes Latin characters with regional scripts, at varying angles, distances, and lighting conditions. A single-pass OCR engine tuned for one character family produces garbled output on mixed-script frames, and a unified multilingual model cannot be independently tuned per family at the accuracy threshold this application requires. The dual-pass architecture runs separate Latin and regional script recognisers in parallel, merging their outputs spatially using bounding box overlap and per-region confidence scoring. Text regions where both passes return a detection use the higher-confidence reading. Regions where only one pass returns a detection use that reading directly. The merge produces a combined text object that handles mixed-script frames without requiring a single model to handle both families simultaneously.

Raw OCR output from field photographs is noisy, unstructured, and mixed with background text that is not part of the data record. Extracting validated fields -- latitude, longitude, application number, beneficiary name -- requires a model that understands the regional layout and format rules that govern how this data appears in stamps. The language model runs and does inference calls, using runtime-reloaded regional configuration templates that define the parsing rules for each state's format. This means compliance teams can update parsing logic for new regional formats without a code deployment. A grounding pass verifies every extracted field against the raw OCR blocks before the field is written to the verdict, eliminating hallucinated values that would pass validation but did not appear in the image.

Confirming that active solar panels are physically installed and correctly positioned requires isolating the panel region within the image and computing spatial coverage metrics. The segmentation model identifies and outlines solar array components, produces coordinate and area outputs for each detected region, and confirms whether the detected configuration meets the presence threshold for that submission type. Open-vocabulary segmentation was chosen over a fixed-class object detector for accessory identification because the set of required accessories varies across deployment configurations and changes over program lifecycles. A model that accepts text prompts at inference time rather than encoding class lists at training time eliminates retraining when a new accessory category is introduced.

Verifying that an operator or beneficiary is physically at the site requires detecting a person in the frame. The presence detection component confirms site visitation without storing biometric. Detection runs and the output is a binary presence signal -- a person is or is not in the frame -- with no persistent biometric record created.

A photograph submitted in one state region and resubmitted in another to claim a second disbursement is indistinguishable from a legitimate submission by visual inspection. Duplicate detection operates against a secure local registry of perceptual image signatures accumulated across all prior submissions. The match runs at submission time, before any downstream processing, and produces a duplicate flag that routes the submission to immediate rejection with evidence. The registry grows with every approved submission.

Running semantic segmentation, dual-pass OCR, pixel forensics, and language model inference simultaneously on workstation-class hardware creates memory pressure that causes out-of-memory crashes on naive parallel implementations. The hybrid execution architecture separates tasks into two tiers: lightweight, non-interfering tasks run concurrently through a thread-pool executor, while high-memory models -- specifically the open-vocabulary segmentation model -- are queued sequentially with explicit GPU cache reclamation and garbage collection between phases. This keeps the full pipeline stable on standard workstations without dedicated VRAM budgets.

Validation patterns, coordinate range rules, application number templates, and language model parsing prompts vary across state programs and change over program lifecycles. Encoding these as static configuration requires a code deployment every time a format changes. The dynamic configuration layer auto-reloads the regional profile at runtime, allowing administrators to modify any validation rule without restarting the service. This means the system can serve multiple state regions from a single deployment and respond to regulatory format changes on the same day they are announced.

Forensic modification scoring applied after dynamic exclusion mask subtraction -- chosen over standard full-frame anomaly detection because benign camera overlays on authentic submissions produce false positives that make full-frame scoring unusable in production

Parallel Latin and regional script recognisers with spatial overlap merging -- chosen over a unified multilingual model because independent per-family tuning is required to reach the accuracy threshold this application demands on mixed-script field photographs

Inference for structured field extraction with runtime-reloaded regional templates – chosen with cloud LLM.

Text-prompt-driven accessory identification at inference time -- chosen over fixed-class object detection because required accessory categories change across program versions and cannot be encoded as a static classifier without retraining cycles

Image signature matching against full prior submission history at submission time -- chosen over metadata-based duplicate checks because a resubmitted photograph carries no metadata that identifies it as a duplicate, only its pixel content does

Concurrent lightweight tasks with sequentially queued high-memory models and explicit cache reclamation -- chosen over a fully parallel execution model because simultaneous segmentation and forensic model loads exceed workstation VRAM budgets and cause out-of-memory crashes

Separate internal telemetry endpoint and filtered public endpoint -- chosen to prevent forensic evidence detail from leaking through downstream enterprise integrations while providing full audit trail access to compliance teams