I build Tīrtha, a verification router for coding agents, at Spanda Works, a small AI company in Rock Hill, South Carolina.
Tīrtha checks an AI's answer before it hands it to you. Tool calls are checked against your schema. Code runs against your tests in a sandbox, and you only get code that passes. Every response says whether it was verified, and when there was nothing to check against, it says that too.
Keep your client, your prompts and your tools. Change one line, your base URL. Free to test: 100 requests a month, no card. After that, each request is billed at the level that served it.
- Director of IT, Savoy ISD, Texas. Ran a school district's technology on a school budget and taught every day. In 2006 our UIL Computer Science team won the Texas state championship.
- Sixteen years at Vanguard in compliance oversight. It teaches one habit for life: do not trust a claim until you can check it.
- Now: building that habit into the tools AI agents use.
| tirtha.ai | The gateway. Free beta, self serve keys from your dashboard. |
| Tīrtha Bench | Every model we test on the same tool-calling tasks: accuracy, speed and cost per right answer, with every raw row kept. |
| Capabilities | What we have measured, generated from the live configuration, including what we have not measured. |
| verification-starter-kit | Free tools for catching an AI system, or the benchmark grading it, reporting a confident number it has not earned. |
| crystal-memory | Agent memory that arrives at the moment of the act, not when you remember to search for it. |
| Report a wrong answer | Found one? Tell us, with or without a key. |
I write about what breaks when you measure your own AI work honestly, including our own mistakes, on dev.to. A few recent ones:
- Four models, one job each: what our own logs say about splitting the work
- Tīrtha Bench: the model data coders actually need
- Our AI judge listens to how sure you sound. Our first number about it was wrong.
- Implementation is where judgements go to become invisible
- Escalating to the better model made 34 answers worse
If you work on AI reliability, I would like to hear from you: support@tirtha.ai.