On your hardware, against your most sensitive corpus, in a network you already defend. That is a large amount of trust to ask for, and the usual answer is a logo wall and a compliance badge. We do not have either yet. Here is the company instead: who runs it, what we wrote ourselves, and what we are still short on.
Every serious data platform of the last decade moved to SaaS, and the AI wave accelerated it. That is a fine trade for most companies. For a bank under supervisory rules, a defence contractor on a controlled network, or a hospital holding patient records, it is not a trade at all. Their answer to can we send this corpus to a frontier lab is no, so the whole category is simply closed to them. The labs will not train inside anyone else’s network, and they will not hand over weights at any price.
ThinkingDBx exists to build the model anyway, on hardware inside their compliance boundary, from a corpus they supply or one we assemble for them. What they get is weights they own outright, with the eval harness and the configs to reproduce the run. Their firewall rules are the control, not our promises. Building for that constraint turned out to produce something a second group wants too, teams with no compliance problem at all who simply want a model that knows their domain. We take that work as well, but it is not what shaped the company.
Committing to that has a consequence most vendors avoid: you cannot outsource the parts that touch the data, and you cannot run someone else’s fine-tuning API behind a project plan. So we wrote the stack. A memory database for the agents, a compiled language where tables and streams are types, a terminal agent, and a 9B mixture-of-experts model pretrained from scratch on our own infrastructure. When a regulated build depends on a vendor, that depth is what you are actually buying. Bonacci Studio, the data platform, came out of the same requirement and is sold separately to the teams who want it.
We are early enough that the founder runs the audit weeks himself. That is a feature while it lasts.
Pretrained the 9B model, wrote the platform, and runs the audit weeks. Most of the engineering posts on the blog are his, including the training write-up and the infrastructure decisions we took the hard way.
ThinkingDBx Pvt. Ltd. was founded in 2025 and is recognised under the Government of India's Startup India programme by the DPIIT. We are self-funded and we run our own infrastructure, down to the Linux boxes this website is served from.
We do not publish a customer logo wall, because we are not going to invent one. The design partner programme is open for exactly this reason: early teams run Studio on real pipelines, meet us monthly, and keep lifetime free access.
We hold no SOC 2 or ISO 27001 attestation. The security page sets out what we do instead and what we will not claim. If certification is a hard gate in your process, we are the wrong vendor this year.
What we can show you is the software running air-gapped on a laptop with the network cable pulled, the configuration file with every endpoint in it, and the case studies on the blog where our own platform did the analysis on public data you can re-run yourself.
No qualification round, no forwarding to an account executive. Bring the task, whatever data you hold, and your compliance boundary if you have one, and you will get a straight answer about whether this fits. Including when the answer is that you should not build a model at all.