AI Engineer & Developer Advocate | Building AI products in public | React, Next.js, GenAI
Joined March 2024
- Tweets 61
- Following 82
- Followers 30
- Likes 32
24 Photos and videos
AI models love "fixing" how our families talk.
I tested @claudeai , @GeminiApp, and @Alibaba_Qwen for Spoken.Kitchen. Most over-correct bilingual slang or crash JSON under stress.
A "Noise Inversion Paradox"—some actually perform better when the text is messy.
Link👇
1
5
7
Jun 9
Market research used to mean 40 tabs and pricing pages that never load.
I did the whole thing in an afternoon — from my terminal, with @nimble_search.
The unlock: every business question becomes a command you can run.
👇Writeup scripts
1
29
Jun 9
Writeup: tinapark.dev/blog/2026-06-07…
Repo raw data: github.com/devtinapark/spoke…
Interactive results: devtinapark.github.io/spoken…
17
Jun 2
Building Spoken Kitchen to save family recipes
I tested @OpenAI GPT-4o vs @DeepgramAI Nova-3 for transcribing English-Korean in a noisy kitchen.
GPT-4o was more accurate.
Nova-3 was faster & cheaper.
For Spoken Kitchen, accuracy won.
spoken.kitchen
#BuildInPublic
42
Tina Park retweeted
Jun 1
We've reset 5-hour and weekly rate limits for all users on Pro and Max plans.
We fixed an issue that caused some Claude Code sessions to spawn excessive parallel subagents, burning through usage faster than expected.
1,112
1,044
20,496
2,548,453
Jun 1
I benchmarked @OpenAI GPT-4o-transcribe vs @DeepgramAI Nova-3 on real Korean/English audio for spoken.kitchen.
🏆 GPT-4o-transcribe
• Better accuracy (CER 0.053 vs 0.077)
• More robust to noise
⚡ Nova-3
• ~2× faster
• ~13% cheaper
Results 👇
2
78
Jun 1
📖 Full write-up:
tinapark.dev/blog/2026-06-01…
💻 Code dataset:
github.com/devtinapark/korea…
31
Jun 1
Another night at Visible Builders OC 🚀
People come for the first time curious.
They come back because everyone is actually building.
No talk. Just shipping.
Grateful for the community.
Let's #buildinpublic
1
4
88
the '5-word prompt triggered a 5h lockout' thing is usually the rolling window already near full, the prompt just tipped it. you never saw it coming because the bar's hidden. claude-meter reads anthropic's usage endpoint so the 5h window is visible before it locks claude-meter.com/r/asq5zj2z
1
1
1
64
May 29
Nothing kills flow state faster than hitting a daily AI rate limit right in the middle of a heavy development sprint. 😩
If you're using Claude Code, a short 5-word prompt can trigger an unexpected 5-hour lockout.
Here is why it happens and how to fix it. 👇 (1/5)
4
107
May 29
🛡️ 1. Token Firewall: Use .claude/settings.json to hard-block tools from crawling heavy dependency paths.
🧠 2. Conceptual Split: Brainstorm on DeepSeek R1 via @OpenRouter ($0). Use CLI only for execution to get @Anthropic's 90% prompt caching discount. (4/5)
1
106
May 29
🧹 3. Flush Context: Run /clear after every sub-task to drop rolling history tax back to zero.
I open-sourced the config framework so you can stop bleeding tokens and stay in your zone.
Full breakdown & repo here:
🔗 tinapark.dev/blog/2026-05-28…
📦 github.com/devtinapark/claud… (5/5)
56
May 26
My Korean mom never writes her recipes down.
So I built an AI that turns her voice into written recipes 💛
Now supporting Korean, Spanish, and Chinese so more families can preserve theirs.
Join waitlist: spoken.kitchen
74
May 18
I vibe-coded my hackathon project in a day. It looked done.
Then I sat down to plan the path to production and realized I'd be unwinding decisions for months.
🧵
1
1
70
May 18
I packaged the interview process as a copy-pasteable prompt.
Paste it into any LLM. It runs a structured interview, pushes back on, and outputs a build spec at the end.
Plan 30–60 min. Link in reply below ↓
What's the failure mode you've hit most building with AI?
1
79