AI Code Review Tool Cuts Per-Run Cost from 50 Cents to 4 Cents with Architecture Overhaul

·2 views

Developer Maneshwar, building an AI code review tool called LiveReview, found that costs reached 50 cents per single run after Gemini File Search's free tier expired, driven largely by a reasoning model being used for retrieval tasks. The core problem was that Gemini 3.5 Flash's reasoning feature billed at $9.00 per million output tokens, meaning the tool was paying a thinking model simply to locate and return relevant text passages. A structural fix — separating the search step so one query could feed multiple draft generations — reduced costs from roughly $0.70 to $0.12 per run, but the underlying issue remained. The team ultimately replaced cloud-based retrieval with a local Chroma vector store using the Qwen3-Embedding-0.6B model, BM25 search, and Reciprocal Rank Fusion, while switching inference to DeepSeek V4 Flash at about one-tenth the token cost. The combined changes brought each run down to approximately 4 cents, a reduction of over 90 percent from the original price.

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)

Git Worktrees Offer a Clean Fix for Running Multiple Coding Agents in Parallel

Developer Andrew Pyle describes how running two AI coding agents simultaneously on the same project caused conflicts when both edited files in a shared directory. His initial instinct to coordinate agents through locks or queues proved flawed, as adding more agents only multiplied the coordination complexity. The cleaner solution, he argues, is eliminating shared state entirely rather than managing access to it. Git's built-in worktree feature allows each agent to operate in its own isolated checkout on a separate branch, all sharing one underlying object store without duplicating full repository history. This approach keeps setup costs low, scales easily to many parallel agents, and simplifies merging once each agent's work is complete.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How accidental robots.txt and sitemap changes silently hurt your search rankings

Two common SEO mistakes — deploying a staging robots.txt that blocks all crawlers and losing sitemap sections after a plugin update — often go unnoticed by site owners because they don't affect the visitor experience. Search engines, however, detect these changes gradually, and the resulting drop in organic traffic may only become visible in reports weeks later. Developers are advised to monitor key files like robots.txt, sitemap.xml, and ads.txt for unexpected changes, and to add automated smoke tests to their deployment pipelines. A sample bash script can flag potential de-indexing signals — such as a broad Disallow directive or a suspiciously small sitemap — immediately after each deploy. For controls embedded in HTML markup or HTTP headers, complementary tools like Google Search Console or scheduled site crawlers are recommended alongside file-change monitors.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How to Monitor Web Page Changes Without Building a Custom Scraper

Tracking updates on web pages — such as registration notices, pricing tables, or government announcements — can be done without writing custom scraping code. The first step is checking whether the page already offers an RSS or Atom feed, which is the simplest and most reliable monitoring method. For server-rendered pages, basic text-monitoring tools work well, while JavaScript-heavy pages may require browser-based monitors like Visualping or changedetection.io. Users can reduce false alerts by targeting specific page sections and avoiding dynamic elements like timestamps or view counters. Monitoring solutions range from free browser extensions and self-hosted scripts to paid hosted services, each offering different trade-offs between cost and technical effort.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

TypeSafe Launches Jev, an AI Model Built to Make Decisions, Not Generate Text

TypeSafe AI has introduced Jev, which it calls a System One model, designed to output structured decisions that software can consume directly rather than generating human-readable prose. Unlike traditional large language models that produce token sequences, Jev works by evaluating predefined decision options and returning typed outputs such as classifications, ranked scores, or calibrated probabilities. The model supports three core decision primitives: selecting from a fixed set of options, scoring on an ordered scale, and returning a yes/no probability for binary decisions. TypeSafe argues that many software tasks — such as routing, sentiment classification, guardrails, and escalation logic — do not require a generated explanation, only a reliable structured answer. The launch raises a broader architectural question about whether AI pipelines need a dedicated, lightweight decision layer separate from reasoning and generative models.

0 comments Read more at DEV Community