Engineer cuts AI inbox costs by limiting LLMs to judgment tasks, not email logistics

·1 views

A software engineer overhauled their email automation workflow by reassigning responsibilities between code and large language models. The core insight was that deterministic tasks — such as routing rules, deduplication, retry logic, and sync state — should be handled by code, while LLMs like GPT-5 or Claude are reserved only for genuinely ambiguous decisions like classifying unclear messages or drafting replies. The engineer noted that repeatedly querying frontier models for repetitive checks is effectively expensive amnesia rather than intelligent automation. Google's Gmail API quota costs — ranging from 2 units for history lookups to 100 for sending — were highlighted as design signals encouraging event-driven, selective fetching over constant full-inbox polling. The revised architecture is described as cheaper, more debuggable, and more resilient than treating an LLM as the central orchestrator of an inbox pipeline.

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)

How Screenshots Caught Four False Positives That Fooled Automated Web Checks

A developer spent an evening auditing 27 third-party product configurators for breakages, only to find that automated tooling incorrectly flagged four fully functional applications as broken. Each false positive shared the same misleading signals: an empty DOM and near-zero page text, which can legitimately occur when a UI launches on a click, renders entirely in WebGL, sits behind a login, or is simply a landing page about the tool. Standard automated checks — including raw HTTP requests, clean browser profiles, and network error filters — failed to distinguish these healthy states from genuine failures. A single screenshot, taken before any measurement, correctly identified all four cases in under a second. The author concludes that automation can only confirm reachability, not functionality, and that human visual inspection remains the most reliable diagnostic step.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

AI in Peer Review Raises Confidentiality Concerns Before Quality Debates

Using AI tools during peer review may breach confidentiality obligations, as manuscripts are unpublished works whose authors never consented to third-party disclosure. Transmitting a manuscript to an external AI service constitutes disclosure regardless of whether the provider stores or trains on the data. Major funders, including the US National Institutes of Health, have banned reviewers from using generative AI to analyse or critique grant applications on these grounds. Journal and conference policies vary widely, with some prohibiting manuscript uploads to external tools while others require disclosure or remain silent. Experts also warn that AI-generated reviews tend to be fluent but lack genuine expert judgement, making it difficult for editors to distinguish them from authentic assessments.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

AI Writing Tools and Paper Mills Are Different Problems Needing Different Fixes

Experts warn that academic integrity debates conflate two distinct issues: researchers using AI to draft or polish their work, and fraudulent paper mills submitting fabricated manuscripts to pad publication records. The first is considered a disclosure matter, while the second is outright fraud that predates generative AI, having previously relied on image manipulation and fabricated data. Policies targeting AI-generated text risk penalizing legitimate assisted writing while failing to catch lightly reworded fraudulent content. Paper mills persist because hiring and promotion systems reward publication counts, creating demand that generative tools made cheaper to supply but did not originate. More effective interventions focus on verifiability — such as requiring open data and code — rather than attempting to detect how content was produced.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How Intended Use Statements Determine If Clinical Software Is a Regulated Device

Whether clinical software qualifies as a regulated medical device depends on what its manufacturer claims it does, not how it is technically built. In the European Union, software is classified as a medical device under MDR (EU) 2017/745 if it is intended for a medical purpose such as diagnosis, treatment, or disease monitoring, with no exemption for clinical decision support tools. In the United States, the 21st Century Cures Act of 2016 added a carve-out under section 520(o) of the FD&C Act that can exclude certain clinical decision support software from the device definition, a provision that has no EU equivalent. Under EU rules, once software qualifies as a device, it is classified using Rule 11 of MDR Annex VIII, ranging from class IIa by default up to class III if its outputs could lead to death or irreversible health harm. Regulatory guidance MDCG 2019-11 remains the key reference document for European qualification and classification analysis, and promotional materials are treated as evidence of a product's intended purpose.

0 comments Read more at DEV Community