Services
By role
Tools & Supplies
My Bookmarks
Example. Asking a model to find a "zebra" in an image when it was only trained on "horses" and "stripes," but never actual zebra images.
See also. CLIP, Foundation Model, Transfer Learning
Example. Semiconductor plants using CV to detect microscopic wafer defects in real time.
See also. Automated detection, Predictive Maintenance
Example. Automatically sorting images and flagging only uncertain ones for human review.
See also. AutoML, Efficiency
Example. DuckDB running in 1.5s on a warm cache vs 2.3s on cold start.
See also. Latency, Caching, RAM
Example. Searching for all images that look like a reference defect image.
See also. Similarity, Query item
Example. Systems with limited customization or potential vendor lock-in vs. self-hosted solutions.
See also. Embedding, Similarity Search
Example. Clustering similar images, dimensionality reduction, anomaly detection.
See also. Clustering, Supervised Learning, Self-supervised Learning
Example. A huge archive of security footage without tags or a searchable database.
See also. Metadata, Search & Retrieval
Example. Fine-tuning a car recognition model to specifically detect factory defects.
See also. Fine-tuning, Foundation Model
Example. A clean dataset with accurate annotations provides a strong training signal.
See also. Loss Function, Gradient, Ground Truth, Noise
Example. Images processed per second or frames per second in video analysis.
See also. Latency, Scalability, Inference
Example. Setting a threshold of 0.95 to identify near-duplicates.
See also. Similarity, Precision/Recall
Example. Patching together multiple custom scripts that eventually break at scale.
See also. Refactoring, Homegrown tools
Example. Using GANs to create training images; simulation data for autonomous vehicles.
See also. Generative Model, Data Augmentation
Example. Image classifiers with labeled photos, object detection with annotated bounding boxes.
See also. Ground Truth, Classification Model, Discriminative Model
Example. A folder structure with consistent naming conventions for "cats" vs "dogs".
See also. Dataset Curation, Data Integrity
Example. Writing a SELECT * FROM images WHERE label = 'cat' query to filter data.
See also. DDL, DML, Relational Database
Example. Finding all variations of a specific defect type.
See also. kNN, Visual Search
Example. Images of the same scene from slightly different angles; products with shared visual characteristics.
See also. Cluster, Similarity
Example. Identical images having similarity of 1; unrelated having 0.2.
See also. Distance, Embedding, Cosine Similarity
Example. Data collection teams using different tools than data science teams.
See also. Collaboration, Workflow automation
Example. Labeling every pixel as "road," "car," "sky," or "building" in a driving scene.
See also. Instance Segmentation, Polygon, IOU
Example. Searching for "a crowded street at night" would find relevant photos without needing specific tags.
See also. Vector Search, Embedding, Natural Language Query
Example. DINO, predicting masked image patches, predicting image rotations.
See also. DINO, Foundation Model, Unsupervised Learning
Example. Selecting multiple mislabeled images to apply bulk remediation.
See also. Remediation, Dataset exploration
Example. Selecting 5-10 clear images of scratches to teach the system what "Scratch" defects look like.
See also. Active Learning, Label Propagation, Few-Shot Learning
Example. Deploying a 7B model on a cloud cluster vs. a smaller model on an edge device.
See also. Edge AI, Inference, Latency
Example. Backing up DuckDB files and metadata snapshots for long-term archival.
See also. Object Storage, Data Lake
Example. Investing in visual inspection to save on scrap costs.
See also. Business Value, Efficiency
Example. Re-labeling an image incorrectly tagged as "truck"; removing duplicates.
See also. Data Cleaning, Annotation
Example. Training robots to navigate spaces, autonomous vehicles learning driving policies.
See also. Edge AI, Active Learning
Example. A policy learns to allocate traffic across servers to minimize latency; an agent learns to play a game by receiving points for winning moves.
See also. Markov Decision Process (MDP); Policy; Reward Function; Exploration vs. Exploitation
Example. Predicting house prices, estimating object angles, measuring defect size.
See also. Prediction
Example. Of all actual defects in the dataset, how many did the model find?
See also. Precision, F1 Score, Sensitivity
Example. Search engines ranking webpages, recommender systems, similarity search in computer vision.
See also. Similarity Search
Example. Loading a large dataset into RAM for fast in-memory processing by DuckDB.
See also. Memory, Cache, Latency
Example. Selecting a reference image to find visually similar content.
See also. Visual Search, Similars
Example. A model deployed on an assembly line to automatically detect defects.
See also. Deployment, Inference
Example. In healthcare, removing patient identifiers; in vision, face blurring or synthetic data augmentation.
See also. Synthetic Data, Data Governance
Unlike generative AI, which creates new content, predictive AI focuses on forecasting events or behaviors.
Example. Forecasting stock prices, predicting equipment failure (predictive maintenance), predicting customer churn.
See also. Generative AI, Discriminative Model
Example. A model predicting an image contains a "dog" with 95% confidence.
See also. Inference, Ground Truth, Confidence Score
Example. Of all images flagged as "defective," how many were actually defective?
See also. Recall, F1 Score, Confusion Matrix
Example. Storing the "source of truth" for metadata and managing vector embeddings.
See also. OLTP, ACID, pgvector
Example. Tracing the exact outline of a car or the irregular shape of a defect.
See also. Semantic Segmentation, Annotation, Mask
Example. The sequence of data inspection -> preprocessing -> training -> evaluation -> deployment.
See also. Workflow, AI Lifecycle
Example. Using PG extensions like pgvector for similarity search.
See also. PostgreSQL, SQL, OLTP
Example. Visual data collected by drones, CCTV, and body cameras for a large organization.
See also. Scalability, Big Data
Example. A photo uploaded to social media vs. the original file.
See also. Deduplication, Data Cleaning
Example. A p95 query latency of 200ms means 95% of queries finish in under 200ms.
See also. Latency, Throughput, SLA
Example. A model trained on near-duplicate images failing to recognize unique images in the wild.
See also. Regularization, Cross-validation
Example. A cat image in a dataset of buildings; images with extreme brightness/darkness.
See also. Anomaly Detection, Data Cleaning
Example. PostgreSQL handling writes and transactional state for metadata and vectors.
See also. OLAP, ACID, MVCC
Example. DuckDB acting as an embedded OLAP engine for exploration workloads.
See also. OLTP, Vectorized Execution
Example. Installing vision systems on factory-owned servers in an air-gapped environment.
See also. Edge AI, Air-gapped
Example. Detecting and locating all cars, pedestrians, and traffic signs in a street scene.
See also. Bounding Box, Localization, Classification
Example. A car in a street scene; a defect on a circuit board.
See also. Object Detection, Instance
Example. Images with poor lighting, over-exposure, or a distracting background.
See also. Data cleaning
Example. A small feed-forward network predicts house prices from tabular features; a deep network recognizes objects in photos.
See also. Deep Learning; Backpropagation; Activation Functions; Convolutional Neural Networks; Transformers
Example. Computing systems that learn hierarchical representations of visual features.
See also. Deep Learning, Layer, Back propagation
Example. Classify support tickets by topic; extract entities (company, product, bug); generate a summary of a meeting transcript.
See also. Transformers; LLMs; Tokenization; Information Extraction
Example. CLIP (vision+text), Visual Question Answering, Video understanding systems.
See also. LLM, Foundation Model, Semantic Search
Example. Supervised vs. Unsupervised Learning.
See also. Supervised Learning, Unsupervised Learning, Reinforcement Learning
Example. Classification vs. Regression vs. Generative models.
See also. Classification, Regression, Generative, Ranking
Example. A product detection model failing as packaging designs change over the year.
See also. Data Drift, Monitoring
Example. An image of a car labeled as "truck"; demographic mislabeling.
See also. Label Noise, Ground Truth, Annotation
Example. Information like scan date, machine ID, product type, and image file paths.
See also. System Metadata, Indexing, Search Attributes
Example. Continuous integration (CI), continuous delivery (CD), model monitoring.
See also. Deployment, Model Ops, AI Lifecycle
Example. Algorithms learning to recognize cats by analyzing thousands of cat photos.
See also. Deep Learning, AI
Typically involves selecting an objective (loss), training on labeled or unlabeled data, and evaluating generalization.
Example. Train a model to predict customer churn from product usage events; or classify images as “defective” vs “ok” from labeled examples.
See also. Deep Learning; Neural Networks; Supervised Learning; Unsupervised Learning
Example. Predicting the coordinates of a person's face or the location of a defect.
See also. Object Detection, Bounding Box, Keypoint Detection
Example. Neurons in one layer connecting to neurons in the next, progressively extracting features from edges to objects.
See also. Neural Network, Deep Learning, CNN architectures
Example. Time from camera capture to defect detection results.
See also. Throughput, Edge AI, Inference
Example. After providing 10 seed examples, the system automatically labels 800 similar images and sends 100 uncertain ones for review.
See also. Semi-Supervised Learning, Active Learning, Confidence Score
Example. This particular image is labeled "Scratch" (the label is the specific assignment to this one image).
See also. Class, Annotation, Instance, Ground Truth
Example. Finding the 5 most similar product images to a query image.
See also. Visual Search, Clustering, Deduplication
Example. Job: "Product-X-2025", Setup: "Standard-Inspection", Recipe: "High-Resolution-Scan".
See also. Recipe, Setup, Traceability, Data Consistency
Example. First iteration: provide seeds and review results. Second iteration: adjust seeds based on results and review again.
See also. Active Learning, Feedback Loop, Model Versioning
Example. Train/test leakage, duplicate detection, mislabeling, data drift.
See also. Issue, Dataset Analysis
Example. A specific train/test leakage occurrence; a specific cluster of duplicates.
See also. Issue Type, Remediation
Example. Distinguishing "car #1" from "car #2" within an image.
See also. Semantic Segmentation, Object Detection, Polygon
Example. Running a defect detection model across a batch of production images.
See also. Prediction, Deployment, Latency
Example. Recognizing that a photo contains a dog, or that a product image shows a specific item from a catalog.
See also. Image Classification, Object Detection, Computer Vision
Example. Flipping an image horizontally, rotating it, or adding noise.
See also. Synthetic Data, Regularization, Overfitting
Example. Calculating average pixel intensity; applying a Laplacian filter to quantify blurriness.
See also. Metadata, Feature Extraction, Data Quality
Example. Sensors and cameras on a factory floor sending data to a monitoring system.
See also. Edge AI, Digital Twin
Example. Humans reviewing the 2% of images the AI is uncertain about.
See also. Active Learning, Annotation
Example. Python scripts for finding duplicates that crash on large datasets.
See also. Technical Debt, Buy vs Build
Example. If an object is bigger than X pixels, classify it as a defect.
See also. Classical ML, Data Preprocessing
Example. A set of 100 manually verified defect images where domain experts have confirmed each label.
See also. Label, Seed, Annotation, Verification
Example. DALL·E generating images from text, ChatGPT writing essays, synthetic data generation.
See also. GAN, Image Augmentation, Synthetic Data
Example. Generating realistic faces; style transfer (photo to painting).
See also. Generative Model, Synthetic Data
Example. DuckDB's create_fts_index supporting keyword searches across captions.
See also. Inverted Index, NLP
Example. GPT-4 for text, CLIP for vision-language, and Stable Diffusion for images.
See also. Transfer Learning, Embedding, Fine-tuning
Example. A foundation model trained on millions of images can be fine-tuned to detect factory defects.
See also. Transfer Learning, Foundation Model, Few-shot Learning
Example. Filtering a cluster for images with brightness above a threshold.
See also. Metadata, Dataset exploration