Developer builds AI agent to automate dog image curation for ML model training

·1 views

A developer building Todogs, a Pokémon GO-style dog breed recognition app for Android and iOS, needed large volumes of clean, correctly labelled dog images to train and improve a TensorFlow model covering 117 breeds. Manually sourcing and filtering images — checking for wrong breed tags, studio shots, cropped dogs, or multi-dog frames — proved too time-consuming to scale. Participating in Google's All Things Agentic Hackathon, the developer built an AI agent using Google Cloud tools including Gemini, Vertex AI, and the Agentspace SDK to automate the entire pipeline. The agent fetches images from online repositories, applies fast Python-based checks for file size and near-duplicate detection via perceptual hashing, then uses Gemini to verify breed accuracy, image sharpness, single-dog presence, and whether the photo resembles a real-world phone snapshot rather than a studio image. The solution aims to compress a process that previously took weeks into a matter of minutes, and could also be used to vet user-submitted photos from the app as future training data.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Interlace.sh Proposes a Single Abstraction to Unify Fragmented Data Pipelines

Modern data engineering stacks typically require multiple separate tools — such as dbt, Airflow, and dlt — each with its own configuration, testing framework, and mental model, creating fragile handoffs between layers. The team behind Interlace argues that these 'seams' between tools are the root cause of silent failures, such as schema drift going undetected or transformations running on stale data. Their solution, Interlace, treats everything — ingestion, transformation, and orchestration — as a single 'model' abstraction, whether written in SQL or Python. Dependencies are inferred automatically from query syntax or function parameters, eliminating the need for manual wiring or separate configuration files. The project draws a parallel to how general-purpose programming unified fragmented workflows through abstractions like functions and package managers, suggesting data engineering is roughly a decade behind on the same trajectory.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Why AI Agent Debugging Requires Structured Tracing Beyond Simple Logs

Simple console.log statements fall short when debugging AI agents that run tools in parallel, handle retries, and use fallback data sources. Flat log outputs record timestamps but fail to capture causal relationships between events, making it unclear which operations depended on which results. Unlike sequential scripts, AI agents have dynamic control flows where a model selects tools at runtime, concurrent strategies run simultaneously, and one agent may delegate work to another. A tree-structured trace model explicitly maps parent-child relationships between operations, making failures and fallback paths far easier to diagnose. Attaching stable identifiers like trace IDs and parent span IDs to each event gives logs a defined contract, enabling proper reconstruction of what happened and why.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

New Laravel Package Auto-Locks Livewire 4 Components Against Client-Side Tampering

A developer named janecodelife has released an open-source Laravel package called 'livewire-secure-properties' that automatically secures public component properties in Livewire 4 from client-side manipulation. The package requires zero configuration and locks all public properties by default, throwing a security violation exception if a client attempts unauthorized modifications. Developers can selectively allow client-side updates on specific properties by marking them with an #[Unlocked] attribute. The package is installable via Composer and supports both single-file and multi-file Livewire 4 component formats. It can also be globally disabled in specific environments, such as local development, through a simple environment variable setting.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Fireworks AI raises $1.5B at $17.5B valuation by making AI models faster and cheaper

Fireworks AI, a startup focused on optimizing AI model performance rather than building models itself, has raised $1.5 billion at a $17.5 billion valuation. The company, backed by Index Ventures, TCV, and Nvidia, now generates over $1 billion in annual revenue — five times more than the previous year. Fireworks processes over 40 trillion tokens daily, nearly triple its volume from a year ago, by making other companies' AI models run faster and at lower cost. The funding highlights a broader industry shift, where inference — the ongoing cost of running AI models for millions of users — now accounts for over 80% of AI hardware budgets. As AI agents multiply the number of model calls per task, the demand for efficient, scalable AI infrastructure is growing rapidly, making performance optimization a significant business in its own right.

0 comments Read more at DEV Community