I spent the past year solving a real problem: AI-translated subtitles are inconsistent—switching between formal and casual, character names changing mid-film, tone jumping around randomly.

Existing tools treat subtitles as isolated sentences. So I built a *multi-agent system* that works like a real translation team: a "director" creates a style guide for the entire film, a "glossary agent" enforces consistent terminology, and a "translator" works with context windows.

It works well. I've processed hundreds of hours of subtitles with it, and the quality beats most SaaS products I've tried.

Honestly, I originally wanted to build a SaaS. But getting a product noticed is hard—marketing, customer acquisition, support. As a solo developer, I don't have the resources of a big company or VC.

Rather than let this project die on my hard drive, I'm open-sourcing it so people who need it can deploy it themselves. At least this way, the time I spent isn't wasted, and it can help some people.

Three specialized agents work in sequence: 1. Director Agent - Reads the entire subtitle file and generates a style guide (tone, formality, addressing strategy). Uses long-context models to analyze the full script.

2. Glossary Agent - Extracts and enforces consistent terminology: character names (with gender inference for pronouns), locations, domain terms. Outputs a Markdown glossary embeddable in prompts.

3. Translator Agent - Uses a sliding window with bidirectional context: - References previous translations (for coherence) - Previews upcoming sentences (to avoid splitting errors) - Strict format validation (preserves timing, IDs, structure) - Batch processing for efficiency

4-6. Reviewer, Polisher, and Timing Adjuster agents are planned for Q1 2026.

Stack: - .NET 10.0 + Microsoft Agents framework - Subtitle Edit (libse) for parsing - Multiple subtitle formats (auto-detection) - Compact format to reduce token usage - Configurable LLM endpoints (OpenAI-compatible, Gemini, etc.) - Model strategy: Long-context for Director/Glossary, cost-effective for translation

Status: [Y]Core pipeline (steps 1-3) is production-ready and open source [Y]Monolingual translation works [Y]Bilingual subtitle support available Steps 4-6 Before Q2 2026 Windows/macOS/Web UI planned for May 2026

Show HN: I built an AI that calls you and practices spoken English with you

Why Big Tech Performance Reviews Aren't Meritocratic

Time Until Someone Points Out This Is Not a Real Study

Agree or Disagree

Dev Logs Without the Noise (2024)

Ruby_LLM-agents: A Rails agent framework for RubyLLM

The secret fast track for animal drugs (2025)

Rise in Sophisticated Dark Patterns Designed to Trick and Trap Consumers (2022)

Change Blindness in UX (2018)

Rust's Standard Library on the GPU

Community Pulse 2025 End of Year Wrap-Up [audio]

Every Enemy from Super Mario 64, 3D Printed [video]

StatechartX – performant state machine runtime written in Go

Show HN: Open-source multi-agent subtitle translator (self-hosted)

MIT's new 'recursive' framework lets LLMs process 10M tokens

I don't like skiing in the shade, so I built a ski trail shade map

Tour website's AI sends visitors to Tasmanian sites that do not exist

198-Bit Constraint Framework: New Physics from First Principles

Trump FCC threatens to enforce equal-time rule on late-night talk shows

NexDock is building a new Windows phone that you can buy in 2026

Elizabeth Holmes asks President Donald Trump to let her out of prison early

Tsfresh

The Art of Craftsmanship (Monozukuri) in the Age of AI

Myth of the Monolithic ERP: Why They Keep Failing [video]

An A.I. Startup Says It Wants to Empower Workers, Not Replace Them

Testosterone went from prostate cancer villain to potential ally

Flashlabs releases the world’s first open-source voice cloning model

Show HN: imessage-data-foundry – Synthetic iMessage Data Generator

Open4D – Open-Source 4D Geometry Processing, Compression and Streaming Library

Palantir CEO: With AI, economies won't need immigration

Show HN: GluonDB – Cursor for Your Database