Developer Builds Production AI Agents for Coffee Shop via Google Cloud Gen AI Academy

·1 views

A developer participating in Google Cloud's Gen AI Academy APAC Edition (Cohort 3) built a multi-agent AI system designed for a coffee shop business, deploying all components on Cloud Run. The project spanned three tracks: a customer-facing chat agent using WebSockets and Google Cloud Buildpacks, a data analysis tool leveraging Gemini with BigQuery, and an operations automation agent with sandboxed command execution. Security was a recurring theme, with the Principle of Least Privilege applied through dedicated service accounts and human-approval checkpoints before any changes to shared records. Upon completing the program, the developer earned a Gen AI Master Certificate and is now preparing for the Cohort 3 Ideathon to build a secure personal journaling application powered by Gemini.

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)

CTEs vs Subqueries: SQL Optimizers Treat Them Almost Identically

A common belief among developers is that Common Table Expressions (CTEs) are inherently faster than subqueries, but this is largely a myth in modern SQL Server and PostgreSQL. Both CTEs and derived tables are typically expanded into the same relational tree by the query optimizer, producing identical execution plans and I/O costs. However, when a CTE is referenced multiple times within a single query, SQL Server may execute its underlying logic more than once or create a Lazy Spool in tempdb, adding overhead. In such cases, using a temporary table with an explicit clustered index is recommended for handling large, multi-million-row datasets. Understanding these internals helps engineers make more informed decisions when designing data pipelines.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer Builds Browser-Based Byte Decoder to Avoid Third-Party Privacy Risks

A developer has shared a lightweight, client-side tool that decodes raw byte sequences into readable text entirely within the browser, eliminating the need for third-party online decoders. The utility is built using JavaScript's native TextDecoder API and Uint8Array, supporting both decimal and hexadecimal input formats. It is designed for use cases such as debugging network streams, parsing custom file formats, and inspecting database buffers. The tool decodes byte sequences using UTF-8 encoding, making it compatible with modern web standards including emojis and multilingual scripts. By processing all data locally in browser memory, the converter avoids the data privacy risks associated with pasting sensitive byte sequences into external services.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

okf-guard Python library scans documents for hidden prompt injection before AI ingestion

A new open-source Python library called okf-guard aims to protect AI agent pipelines from indirect prompt injection attacks embedded in source documents. The tool targets a specific vulnerability where hidden text in common file formats — such as PDFs, Word documents, spreadsheets, and presentations — gets extracted alongside legitimate content and fed to AI systems without distinction. okf-guard supports six file formats and runs two independent checks on each scan: one for hidden content and one for injection-style phrasing, flagging either finding separately. The library is particularly relevant to Google's recently published Open Knowledge Format (OKF), which has AI agents read markdown files directly with no intermediate processing layer. All detection is rule-based and fully deterministic, with no LLM dependency or network calls, making its behavior reproducible and auditable.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Tutorial Proposes Consent-Based Memory Design for AI Voice Companions

A software design tutorial published on DEV Community argues that AI voice companions should treat user memory as a consent ledger rather than a passive prompt history. The approach requires the AI model to propose a typed fact, which the user must explicitly confirm, correct, or reject before it can be stored. Only confirmed records are permitted to enter future prompts sent to a large language model. The design restricts storable data to a small set of bounded slots—such as preferred name, music genre, and chat style—deliberately excluding free-form instructions or sensitive personal data. The tutorial demonstrates the implementation in TypeScript, integrated with Tencent RTC's Conversational AI platform, emphasizing that the application, not the model, should control what becomes durable memory.

0 comments Read more at DEV Community