Just two characters but still, I had to laugh a bit seeing all the complaints in here about the Chinese language model replying in Chinese. US big tech models can do it too!
I'll include a link at the end but this is not primarily intended to be a promotional piece. I see people occasionally asking after what others are doing in here, and I thought I'd give a breakdown of how I took something from concept to production using DeepSeek in a little over a month.
My initial goal was twofold: learn to use multi-agent AI workflows and see if I could build something real with DeepSeek. I'd read Yegge's epic Gas Town post earlier this year and he was right - I wasn't ready for the type of massively parallel agentic workflows he has implemented. So I thought I'd start with something small that I could scale up on, and use DeepSeek to not break the bank. 99% of the code was written with some combination of DS v4 Pro and Flash. I have spent a total of ~$12 on tokens. At one point I had 8 agents running in parallel, that was cool.
I'm a software developer with 20+ years of experience, so I'm opinionated about code and generally know what I'm doing. I wrote exactly 0 lines of code myself, but I have read pieces of the code and pushed the agents to do things certain ways. I did the initial planning and ideation with my $20/mo Claude subscription. I also had Claude help author the AGENTS.md file, individual agent files, and set up a devcontainer with OpenCode. I continued to consult with Sonnet on and off through the process.
I wanted a devcontainer so I could let the agents run free without prompting me for anything, with no risk of them accessing any credentials/secrets. This turned out to be a good workflow for me personally. Source control is there for disaster recovery, and only I could push to github. Same with all the services: I spent a lot of time doing manual deployment to runpod and fly.io.
Within the source tree, DS v4 Pro did all the planning. Everything was delegated to a flash agent for implementation and reviewed by pro. The project has about 10k lines of code and 800 unit tests, all written in python -- the tests are really key to ensuring things worked without a close examination of everything the agents were doing. So a typical cycle would be plan - implement - test - review - fix - commit.
As to the product itself, it's a pretty straightforward ebook to audiobook TTS application. My niche is in pay as you go, length-based pricing, I'm not trying to compete with ElevenLabs. I'm using the Kokoro model open weights deployed on runpod - GPU time is relatively cheap, and Kokoro weights are not huge.
Getting epub TOC parsing to map to m4b chapters was probably the single hardest part, and the one most likely to still need revision. Epub rendering is pretty straightforward. Extracting text into chapters consistently without duplicating or eliding text required quite a bit more work. I don't think that DS did any better or worse than I would have done in the end. I did take an active hand in correcting its original approach of following the epub spec to lay out the text according to the spine rather than the TOC, and backing the TOC out of the spine. It implemented a bunch of really questionable heuristics and did things a junior dev would do, and the reviewer just waved them on through. The more the code grew unwieldy, I was able to get pro to discuss concerns and come up with a better plan. Having some in-depth conversations about design resulted in something that was pretty good.
Once I asked the agent to change course and it reverted a ton of changes already in progress. A git stash would have been preferred. When I challenged the agent on what the hell it was thinking, it said it got embarrassed. This was possibly my favorite moment working with DeepSeek as an AI. You do have to keep course correcting at every step.
Towards the end I set up Claude Code in the same container and used some tokens from my Claude sub to have Opus 4.8 review everything. It was impressed at the overall quality. It did make some good suggestions for test improvements but it didn't really flag anything that DeepSeek had missed or done poorly. I don't feel like a frontier model would have made an appreciable difference in my workflow.
$12 for about 750k tokens in total. Not bad. If you're interested in seeing what it produced, ebookaloud.com is the live site.