1. ffmpeg -i course.mp4 -vn audio.wav # Extract audio
2. asr_result = gateway.chat(ASR_MODEL, audio) # Get SRT with sentence-level timestamps
3. Split text into 5-minute windows, preserving the timestamp mapping
4. segments = gateway.chat(LONG_CTX_MODEL, prompt_with_timestamps)

Ask the model to output: [{start, end, title, summary}], constrained by JSON Schema

1. Title polishing + course-level description = gateway.chat(LIGHT_MODEL, ...)

Use a lightweight model, costing about 1/10 of the flagship

1. Human spot-check: randomly pick 3 split points to verify (about 2 minutes)
2. Output: chapters.json + subtitle.srt → push to CMS

Three things to watch when implementing: first, the prompt must explicitly state "use only the provided timestamps; do not invent times on your own"; second, use JSON Schema or function calling to constrain the output format, to avoid burning tokens on retries after parse failures; third, process long videos in segments and then merge chapters globally, to prevent exceeding the context limit.

## 4. Why a Unified AI API Gateway Saves Maintenance Costs

This is something many small teams underestimate. If ASR, segmentation, and summarization each connect directly to three different vendors, you end up with three SDKs, three sets of auth logic, three error-code systems, and three bills—and every time any vendor changes pricing or API versions, your pipeline gains another failure point. The code that sets off midnight alerts three months after launch is often exactly this kind of "three-front war" integration code.

The benefits of a unified gateway are very concrete:

- **Authenticate once, call everywhere.** Model upgrades and vendor switches only require changing gateway configuration—zero changes to business code. When we switched our summarization model from one vendor to another, the only change was one line: the model name.
- **Uniform formats.** Different vendors have different streaming responses and error structures; the gateway smooths them out, so retry and fallback logic only needs to be written once.
- **Observable usage.** How many tokens a video consumed, which step is the most expensive—all visible in a single dashboard. This is a prerequisite for cost control and tiered model configuration; without a unified entry point, you can't even piece this data together.

## 5. The Bottom Line

For a 20-video course package: before the transformation, about 40 hours of manual labor (roughly ¥3,200 at ¥80/hour); after, about 3 hours of labor plus under ¥10 in API costs. Even if you only process two course packages a month, the time saved per year approaches 1,000 hours—for a one- or two-person team, that's essentially an extra full-time hire, and this "hire" never degrades from repetitive work.

Run the pipeline end-to-end on a single video first, then scale to parallel batches—that's the most reliable rollout rhythm for this scenario. If you're ready to get started, register an account on a unified gateway and consolidate all your ASR and LLM calls in one place: https://api.thistoken.ai/register

---

Ready to try it yourself? Sign up at https://api.thistoken.ai/register to get your API key and start building.