FrameMill Uses Google Gemini AI to Turn Long Game Streams Into Editable Video Cuts

·1 views

A developer built FrameMill, an AI-powered tool that converts lengthy game stream recordings into editable video projects, created as an entry for the Gemini All Agentic AI Hackathon. The tool uses Google's Gemini AI alongside a Go-based media processor to detect gameplay events and assemble a first-cut timeline, rather than delivering a finished, unalterable highlight reel. Unlike existing clipping tools, FrameMill gives creators a transparent, revisable Timeline as its core output, letting them inspect, modify, and re-render the edit according to their own creative goals. The system currently supports Call of Duty: Warzone DMZ and is built with a stack that includes FFmpeg, Vue, TypeScript, Google ADK agents, and a cloud architecture using Cloud Run and Firestore. A key technical challenge was efficiently processing hours of footage, which the team addressed by decoding video once and sharing frame data across multiple detectors rather than having each component read the file independently.

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)

Developer Builds Free AI Agent Fleet to Help Farmers Detect Crop Problems

A developer created Fieldhand, a free AI-powered multi-agent system designed to help farmers identify and respond to crop health issues, built during the All Things Agentic Hackathon. The system runs automatically at 4am via cloud scheduling, using a fleet of specialized agents — Diagnostician, Skeptic, Agronomist, and Operations — to analyze satellite data from Sentinel-2. A key design principle is that the Skeptic agent actively tries to disprove diagnoses rather than confirm them, using field history and neighboring comparisons to filter out false alarms. In one test, the system correctly dismissed a catastrophic-looking alert on a Palouse field by recognizing it was a historically bare, post-harvest plot each August. Any action involving costs requires human approval, enforced through a model-free policy layer to prevent unauthorized spending even if an agent is compromised.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer finds same wrong number in six files but cannot trace its origin

A software developer discovered that the figure 1,803 — cited as an element count in a planning ledger — could not be reproduced by any mechanical count of the actual file, with an automated script returning 1,774 checked rows instead. A search across the project's planning tree found the identical incorrect value repeated in six separate documents, suggesting the number was copied from file to file rather than independently measured each time. The developer could not identify where the figure originally came from, and the published correction honestly states the number does not reproduce from the source rather than claiming a root cause was found. Rather than fixing the files silently, the developer added inline dated notes to each of the six occurrences, leaving the wrong number visible alongside the correction to preserve auditability. Verbatim transcripts and unrelated grep matches, such as an arXiv identifier containing the digits, were deliberately left untouched to avoid introducing new inaccuracies.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Blog Comment Exposes AI Pipeline Flaw, Prompting Three-Layer Audit Fix

A developer team discovered a critical vulnerability in their AI review pipeline after a reader's blog comment challenged the robustness of their provenance-based fix. The original system recorded which source artifact each AI output was derived from, but a commenter pointed out this could still be gamed by the same type of hallucination failure. In response, the team implemented three safeguards: read receipts at hand-off, SHA256 fingerprints at queue entry, and re-derivable verbatim citations audited nightly. An audit of 2,038 reviews found four contaminated entries — roughly 0.2% — which had already caused around 470 wasted AI generations before detection. The team also acknowledged a prior coding oversight that allowed queue entries without a hash to bypass the audit entirely, a gap that has since been closed.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Google Search Console Rolls Out Global AI Performance Reports for All Sites

Google has made dedicated generative AI performance reports in Search Console available to websites worldwide as of August 31, 2026, after an initial limited rollout on June 3, 2026. The reports give site owners a separate view of how often their URLs appear in AI-driven features such as AI Overviews, AI Mode, and Discover. Key metrics include impressions, pages, countries, devices, and date-based breakdowns, helping teams pinpoint which content gains AI-feature visibility. Crucially, the new reports are distinct from the existing Performance report and do not merge AI appearances with traditional search metrics. Publishers also have a separate control within Search Console to manage whether their content is eligible to appear in these generative AI features.

0 comments Read more at DEV Community