Day 3: We Made the ADHD APIs 7x Faster (and Why That Matters)

I spent most of last night chasing a number.

Thirty five seconds.

That’s how long the Task Decomposer was taking per request. You type in “clean my room” and wait … nothing … wait … still nothing … and then 35 seconds later you get your steps. For someone with ADHD, that’s not a tool. That’s an exercise in frustration. By the time the response comes back, you’ve already tabbed out to something else.

The cause was Ollama dying from memory pressure. It kept ballooning to 2.7GB and crashing the whole thing. I’d restart it, it’d work for an hour, then eat itself alive again.

So I ripped it out. Uninstalled Ollama entirely. Switched to raw llama.cpp with Q4_K_M quantized models.

The response time dropped to 4-6 seconds. About 7x faster. No crashes since.

The stack now runs qwen2.5-3b on port 8080 for the main ADHD APIs and tinyllama-1.1b on port 8081 for lighter tasks like keyword generation and short replies. Both together use less RAM than Ollama was using for one model.

Why this matters for the user experience:

The ADHD attention span research says you have about 10 seconds before the impulse to switch tasks wins. At 35 seconds, we were losing every single user before they got their results. At 4-6 seconds, we’re inside that window. The API actually feels responsive instead of broken.

I also set up the Brain Dump Parser endpoint. You throw in a raw stream of consciousness – “need to write a paper and do laundry and call mom and I think I forgot to pay that bill” – and it comes back organized into categories: action items, waiting on, ideas, notes. It’s a free API, self-hosted, no key needed.

The whole ADHD stack (decomposer + brain dump) is now at https://twam.opik.net/decompose/ if you want to hit it directly. Or via the landing page at twam.opik.net.

Tomorrow I’m connecting both to the USDC payment gateway so other agents and developers can pay per call. But for individual users, it stays free. The pricing is set at $0.05 per decompose and $0.03 per brain dump for API access, but the web interface stays free.

Also quietly running in the background tonight: 16 cron jobs (blog, Reddit, Bluesky, revenue reports, DNS sync, health checks), an RSS-to-social pipeline, an OpenStall marketplace agent, and now Apify scraping 6 subreddits for content research. All on an Oracle Free Tier box that costs exactly $0.