Obz AI: Explainable, Observable Computer Vision

• Obz AI is a full-stack software ecosystem for computer vision that integrates model explainability, feature extraction, anomaly detection, and monitoring.
• The platform employs XAI methods like Captum, attention maps, and CDAM to generate and log attribution maps alongside predictions for audit and debugging.
• Its modular architecture, comprising a Python client, FastAPI backend, SQL and S3 storage, and a React dashboard, ensures secure and project-isolated observability.

Searching arXiv for the target paper and closely related context. Obz AI is a full-stack software ecosystem for computer vision that is purpose-built to make deep models both explainable and observable in real-world deployments. It was developed to close a practical gap between maturing explainable AI methods for convolutional neural networks and vision transformers, on one hand, and the limited integration of those methods into MLOps and observability workflows, on the other. In Obz AI, inputs, predictions, explanations, features, and anomaly signals are captured together, stored securely, and visualized coherently for audits, debugging, and responsible deployment, with a pipeline spanning a Python client library, backend services, storage, and an analytics dashboard (Chung et al., 25 Aug 2025).

1. Concept and scope

Obz AI targets modern deep computer-vision models, specifically CNNs and ViTs, and has been demonstrated on image classification tasks. Its central contribution is to make explanations first-class monitoring artifacts rather than isolated post hoc visualizations. In practice, this means that attribution maps are logged and stored alongside predictions, extracted features, and outlier flags under project-aware schemas, so that explanation data can be queried, audited, and compared across runs or time windows (Chung et al., 25 Aug 2025).

This design situates explainability within knowledge management and monitoring rather than treating it as a separate analysis step. A machine learning engineer can incorporate XAI methodologies, extract features for outlier detection, and continuously monitor models in real time through a single ecosystem. The system was developed from scratch to support easy migration between cloud and on-prem deployments, which is particularly relevant for privacy- and compliance-sensitive domains such as medical imaging or surveillance.

A common misunderstanding is to reduce Obz AI to an explanation viewer. The software is broader in scope: it binds explainability, feature inspection, anomaly detection, authenticated logging, storage isolation, project administration, and dashboard-based analysis into one operational workflow. Another possible misconception is that it already covers the full range of CV tasks. The current release focuses on classification for CNNs and ViTs; segmentation, synthesis or generative vision, and broader CV tasks are part of the roadmap rather than the demonstrated scope.

2. Architecture and data flow

Obz AI consists of four principal components that are sequentially connected into a monitoring pipeline: a Python/PyTorch client library installed alongside the user’s model, a backend service implemented with FastAPI, a data management layer using relational SQL and S3-compatible object storage, and a frontend analytics dashboard built with TypeScript and React (Chung et al., 25 Aug 2025).

Component	Role
Python/PyTorch client library	XAI tools, data inspection or feature extraction, and logging client
Backend service (FastAPI)	Authenticated ingest, orchestration, retrieval, and mediation between clients, storage, and dashboard
Data management layer	PostgreSQL via Supabase for structured metadata and S3-compatible storage for images and explanation artifacts
Frontend analytics dashboard	Real-time monitoring, visualization, project administration, and data exploration

During inference, the Python library computes model outputs such as class probabilities, generates attribution maps at image resolution, extracts features for anomaly detection, packages these elements as a log record, and sends the record to the backend with project and user context. The backend authenticates the request, writes structured metadata to SQL tables including users, API tokens, projects, logs, and reference features, and uploads large artifacts such as images, heatmaps, and embeddings into the project’s private object-storage bucket. The dashboard then reads via backend APIs to provide time-series views, outlier panels, per-sample drilldowns, and explanation overlays.

The architecture is organized around project isolation and scoped access. Each project uses an isolated bucket, access is scoped via API tokens, and users can delete data or logs directly from the dashboard. This coupling of deployment flexibility, isolated storage, and deletion controls makes governance and audit workflows part of the platform rather than external operational add-ons.

3. Explainability methods and explanation evaluation

The XAI layer is built around a general-purpose XAITool abstraction. Obz AI computes per-image or batch attribution maps returned at the same spatial resolution as the input, which is important for side-by-side visual comparison and overlay analysis in vision workflows. Explanations can be computed for the top predicted class or for multiple target classes, and they can be saved as images or NumPy arrays for downstream analysis (Chung et al., 25 Aug 2025).

For CNNs, the platform exposes Captum methods, including Integrated Gradients and SmoothGrad, as well as other gradient-based saliency approaches. For ViTs, it implements multiple attention-map variants and class-discriminative attention maps, or CDAM, which align attention to downstream task signals and can target specific outputs or latent concepts. Attention maps visualize self-attention patterns across transformer blocks, while CDAM uses gradient-based scaling of attention by relevance to a specific classifier-head output, with smooth and integrated variants intended to improve stability and compactness.

Obz AI does not treat explanation generation as sufficient in itself. The EvalTool base class supports quantitative estimation of explanation quality through two metrics: fidelity and compactness. Fidelity measures how well highlighted regions align with the model’s predictive behavior under systematic perturbations, including perturbation curves, while compactness reflects how concentrated and human-interpretable the maps are, for example through fewer and tighter salient regions. This enables repeatable assessment of explanations rather than reliance on one-off visual inspection.

An important technical clarification concerns concept-level interpretation. While concept-based explanations such as TCAV are referenced in the literature discussed by the paper, the demonstrated concept sensitivity in Obz AI is achieved through CDAM’s latent-space mechanism rather than through a dedicated TCAV pipeline. This distinction matters because it locates the current implementation within attention- and gradient-based vision explainability rather than within a separate concept-attribution framework.

4. Feature extraction, anomaly detection, and observability

Obz AI complements explanation logging with feature extraction and outlier detection through the Extractor classes. Two pipelines are emphasized. The first uses first-order statistics inspired by radiomics, such as mean and variance across the image, modeled via Gaussian Mixture Models. The second uses semantic embeddings from visual-LLMs such as CLIP, obtained through Hugging Face AutoModel, with embeddings modeled by principal component analysis and reconstruction loss using the top $r$ principal components used to flag outliers (Chung et al., 25 Aug 2025).

In both cases, reference models are fit on a curated dataset and then applied to new images. Substantial deviation indicates potential anomalies in content or distribution. For first-order radiomic features, deviations from the learned mixture can indicate unusual intensity distributions or textures. For CLIP or related VLM embeddings, reconstruction loss identifies samples whose semantics differ from the reference set. The paper presents this as a modular design in which teams can swap detectors or VLMs depending on the application.

These extracted features are not stored separately from predictions and explanations; they are part of the same explanation-aware logging schema. The result is continuous, real-time observability of both performance-side outputs and explanation-side artifacts. The dashboard supports time-based views of feature distributions and outlier counts, per-sample inspection of predictions, explanations, and anomaly signals together, and overlays that make it possible to inspect where the model focuses and whether that focus is stable across samples or time.

The architecture also supports thresholding on outlier detectors, including reconstruction loss, to flag anomalies, although the current release emphasizes visualization and drilldown more than elaborate alerting policies. Likewise, the platform is designed to complement traditional performance tracking with explanation and distributional monitoring, but the paper explicitly notes that it focuses on real-time observability and explanation evaluation rather than enumerating task-specific metrics such as IoU, mIoU, or mAP.

5. Workflow, dashboard operations, and use cases

The integration workflow for CNNs and ViTs is organized into four phases: installation and initialization, inference instrumentation, logging, and monitoring. A model developer installs the Python package obzai, obtains an API token, initializes ObzClient with token and project metadata, wraps the model’s predict() or forward() step to collect inputs and outputs, configures XAITool and DataInspector, packages inputs, predictions, attribution maps, feature vectors, and outlier flags, uploads reference features or models such as PCA or GMM parameters, and then uses the dashboard to visualize trends, outlier rates, and per-sample explanations (Chung et al., 25 Aug 2025).

The frontend dashboard is tailored to CV workflows. It includes time-series panels for features with counts of total samples and outliers in the chosen range, sample drilldowns for feature distributions relative to reference sets, explainability views showing side-by-side inputs and XAI heatmaps, overlays with transparency controls, project switching, user and token management, data deletion, and a Data Explorer for filtering and CSV export of sample subsets. The backend service exposes authenticated endpoints for ingest, reference-feature upload, and project management, and links logs to users and projects.

Two use cases are described. In an ImageNet natural-image setting described as surveillance-like, a pre-trained multiclass ViT was instrumented with Obz AI using CLIP-based PCA for outlier detection and XAI methods including Attention Maps, CDAM, and gradient-based saliency. In the dashboard, side-by-side explanations highlighted class-relevant regions; one example reports a predicted “golf ball” at 98.6% probability with positive importance inside the ball and neutral or negative scores outside. In a medical-imaging use case based on LIDC-IDRI lung CT, a DINO-based binary classifier predicting malignancy was paired with a GMM over 16 first-order features. Time-series plots showed selected features and outlier counts over configurable ranges, while per-sample panels compared feature values to reference distributions and displayed CDAM, attention, and saliency maps for targeted inspection and audit.

These examples illustrate the intended operational role of the platform. Obz AI is not presented merely as an offline XAI toolkit; it is positioned as a production-side environment for continuous monitoring, retrospective analysis, and explanation-aware audits.

6. Positioning, governance, limitations, and future directions

Obz AI is explicitly compared with observability tools such as WhyLabs, Evidently, Fiddler AI, Arthur, and Roboflow. The paper’s claim is not that those platforms are absent from model monitoring, but that they either focus on non-vision data or provide limited XAI capabilities for vision models. Obz AI’s stated distinctive contribution is the deep integration of ViT-specific attention and CDAM methods with feature-based outlier detection, including radiomics-style first-order features and CLIP embeddings with PCA, within a full-stack monitoring and knowledge-management framework tailored to computer vision (Chung et al., 25 Aug 2025).

Governance and responsible-AI considerations are embedded in the design. Hybrid storage using SQL plus object storage is intended to accommodate large volumes of logs and artifacts; service-oriented components are described as horizontally scalable; batch logging and high-throughput ingestion are supported for production pipelines. At the same time, XAI and feature extraction add inference-side cost, and the platform mitigates that overhead through modular configuration, including lighter XAI methods, batch processing, sampling, and offloading storage to object stores. Privacy and security controls include API-token authentication, per-project isolation, user-scoped access, and dashboard controls for deletion and export.

The limitations are clearly stated. The current release focuses on classification tasks for CNNs and ViTs. Planned extensions include segmentation, generative or synthesis workflows, broader CV tasks, additional XAI methods, and conventional ML integrations. The paper also notes that alerting policies and expanded performance metric tracking would strengthen production observability. This suggests that Obz AI should be understood as an operational bridge between explainability research and deployment practice rather than as a completed, task-universal observability stack.

A final misconception worth avoiding is that the platform equates explainability with responsible deployment by itself. The paper’s formulation is narrower and more technical: by making model behavior inspectable through logged explanations, features, and anomaly signals, Obz AI promotes transparency, fairness, accountability, and responsible deployment. A plausible implication is that its main contribution is infrastructural: it turns explanation outputs into queryable, project-scoped artifacts that can participate in the same audit, debugging, and monitoring workflows as predictions and metadata.