Building Intelligent Engines That Ship.

Computational Latency

Standard Python and Pandas operations were computationally prohibitive (~58s per batch) making granular analysis impossible.

This forced a reliance on weak "static aggregates" (e.g., total spend) that blind the model to critical behavioral trends like delinquency streaks.

Dense User-Time Matrix

Raw transaction logs were pivoted into a dense, cache-friendly NumPy matrix (User × Time) structure.

This specific layout is optimized for Numba iteration, allowing for efficient sequential access to historical payment behaviors.

JIT Compilation & Robust Validation

Applied the @njit decorator to compile custom loops into optimized machine code, strictly reinforced by pytest for logic and PSI for stability.

This unlocked complex stateful features like late streaks and velocity slopes, ensuring mathematical correctness and drift resistance (PSI < 0.1).

Business Impact

Achieved a paradigm-shifting execution speedup and the inclusion of these new behavioral features directly resulted in a +4.2% AUC uplift

Minimized cloud compute overhead by eliminating high-latency processing bottlenecks. Proactively identifies high-risk profiles to mitigate potential bad debt exposure before it occurs.

Invisible Infrastructure

Manual mapping is costly, while standard models fail to detect thin cables and heavily occluded poles in chaotic street environments.

Severe class imbalance (e.g., rare providers) and environmental noise rendered off-the-shelf detection models ineffective.

Hybrid Dataset Strategy

Constructed a proprietary dataset from scratch via manual field collection of street-level infrastructure imagery.

Curated and annotated specifically for this task, then augmented with synthetic data to address severe class imbalance.

High-Res YOLOv8 Engineering

Optimized YOLOv8 training with 1248px resolution scaling to recover fine-grained features like fiber optic cables.

Implemented a custom inference engine with EXIF extraction and geospatial mapping (BPS shapefiles) to link detections to specific sub-districts.

Technical Result

Achieved a robust 0.763 mAP50 (up from 0.52 baseline) and successfully resurrected failing minority classes (~0% → ~50%).

Delivered a deployment-ready tool that automates the conversion of raw street photos into mapped competitive intelligence.

Fine-Grained Visual Ambiguity

Distinguishing between visually similar traditional dishes (high inter-class similarity) was difficult due to a limited dataset of only ~4,200 labeled images.

Standard CNNs struggled to generalize, risking overfitting on the small, proprietary training set.

Limited Proprietary Dataset

Worked with a restricted, proprietary dataset of 15 food classes, requiring aggressive augmentation strategies to prevent data starvation.

Due to data scarcity, the project relied heavily on generating synthetic training signals via pseudo-labeling.

Semi-Supervised Teacher-Student Ensemble

Implemented a Teacher-Student framework where high-confidence predictions (>0.86) from a ConvNeXt Teacher were used to train robust Student models.

Combined the local texture bias of ConvNeXtV2 with the global context of SwinV2 using a Geometric Mean Ensemble to calibrate probabilities

Top-Tier Leaderboard Performance

Proven robustness with Test Time Augmentation (TTA), delivering high-confidence predictions even on unaligned or noisy test images.