I'll be honest up front: I'm not a salesman. Everyone has always told me that, and lately even the AI I build with tells me that. So this is not a pitch. This is me asking for help.
I didn't set out to create any of this. It started because I kept having context window issues while building something else entirely, and I went down a rabbit hole. The rabbit hole kept going. Somewhere down there I got tired of watching my AI coding agent send every single request to the most expensive model available, whether the task needed it or not, and I started building a fix for myself.
What's been embarrassing is I thought I had one story. Then the testing led me to another one. Then another. My website has changed multiple times, which has been a lesson learned in public, the hard way. The good news is there are no clients yet, so that's good :)
Where the testing has left me now is this:
As a gateway I am able to send routine coding requests to lower cost models that get verified before you ever see them. The hard ones escalate to a frontier model. On my benchmark runs that was about 1 in 27 requests. The full setup measured around 95 percent on HumanEval+ (n=164) at roughly 8x lower cost per request. I keep re-running these numbers because honestly I didn't quite believe them either. So far they keep holding.
So based on those runs, that works out to making a Fable 5 subscription last up to 27x longer, because most of the work doesn't need that level of cost to be accurate.
What I don't know yet is whether it holds up on YOUR work. Real repos, real deadlines, weird edge cases. That's exactly what I need testers for.
Security, because you should ask: encrypted at rest, TLS in transit, and your code is never training data. I verified the upstream providers do not train on it, and the routine tier runs on open models I host myself. Local models, privacy, security, and a cache that bends the cost curve down even lower over time. That is my goal, my promise, and what I want help testing and proving.
The deal: it's pre-beta, it's free while it is, and I will keep publishing the real numbers whether they flatter me or not. I feel like I have something special here, something people have been asking for. I just need people smarter than me to hammer on it and tell me where it breaks.
Sign up on the waitlist at tirtha.ai, reply here, or message me and I'll add you to the beta team directly.
Top comments (0)