THINKINGDBX · SOVEREIGN AI FOR REGULATED DATA

Your data cannot leave.
Your AI can still be good.

Bonacci Foundry trains a domain model inside your perimeter and hands you weights you own outright. You do not need a corpus to start: a few hundred real examples is enough, and we build the rest. Start with a one week audit that will tell you whether you need a model at all.

One week corpus audit, free for now · model builds $15k to $120k · air-gap capable · weights delivered, not licensed

train.internal.acme-bank.corp no egress
Run · acme_9b
corpus
tokenizer
pretrain
post-train
eval
Corpus
core_banking
ops_manuals
pii redacted
step
18,400/24,000
loss
1.847
tokens
93.5B/122B
throughput
47.3ktok/s
lossstep 018,400
14:02:11 ckpt saved runs/acme_9b/step-18400
14:02:11 eval held-out acc 0.94 on your ground truth
14:02:12 egress attempts 0 · weights stay in /mnt/models
on delivery acme_9b.safetensors acme_9b.gguf eval/ yours to keep, yours to serve →
Runs on your infrastructure Apache Spark Apache Kafka Docker Ollama vLLM PostgreSQL
What goes in

“We do not have enough data” is the reason most teams never call. It is usually wrong.

Nine times out of ten that is a statement about volume. The domain knowledge is already in your building, and volume is the part we can solve for you. There are three ways in, and only the last one assumes you arrive with a dataset.

  • 1

    You bring a few hundred examples

    A few hundred real cases is enough to characterise a domain. Foundry expands that seed set into training volume synthetically, holding to its distribution so the generated data stays inside your world. It is then decontaminated against the eval suite, so the scores at the end still mean something.

  • 2

    You bring nothing but the domain

    We curate the corpus ourselves from public sources, licence filtered, and build against an evaluation suite we agree with you up front. This is how our own models were trained, including the finance model on Hugging Face.

  • 3

    You bring the corpus

    Your documents, tickets, filings, logs, code. We measure duplication, PII and template collapse before a GPU hour is spent, and tell you in writing if what you have is too thin to justify the build.

  foundry corpus plan acme_9b.yaml
# seed set you actually handed us
seed examples        412 real
  deduplicated         397
  quality filtered     361

# volume we generate from it
synthetic expansion  361 → 84,000
  grounded in seed distribution
  eval contamination   clean

# corpus we curate, when you have none
public domain crawl  11.4B tokens
  licence filtered     8.9B

total corpus         9.1B tokens
# your 412 examples set the target.
# they are not the training set.
How Foundry works

We broke three of our own models learning this.

A model build is mostly irreversible. By the time a bad configuration shows up in the weights, the compute is already spent and the only fix is to run the whole thing again. So the decisions that cannot be undone now get checked first, and the checks came out of runs we paid for ourselves.

  • 1

    Scoped before it is sold

    One week, and we are not charging for it at the moment. We audit your data, measure tokenizer fit on your corpus, and train small models on your actual corpus to see where the curve goes. Most people estimate this from document counts. We measure it, and the report tells you what 9B buys you over a fine-tune in numbers from your own data.

  • 2

    Gated while it runs

    Corpus checks measure duplication, PII and template collapse before compute is spent. Configuration checks block the specific mistakes that cost us those three models. Training then records a second curve alongside the loss, because this class of failure does not show up in the loss at all.

  • 3

    Validated before it ships

    Before anything ships we open the weights up and look inside them: dead embedding rows, layer contribution, special-token health, then a smoke test on the serving stack you will actually run. You get weights you own outright, the eval harness, and the configs to reproduce the run.

  foundry precheck acme_9b.yaml
# checked before a single GPU hour is billed

P1  weight decay excludes norm gains   BLOCK
     exclusion list is empty: every
     parameter decays, including the
     RMSNorm gains. invisible in loss.

P2  checkpoint interval fits tenure    BLOCK
     10000 steps on spot capacity:
     every preemption restarts at zero

P15 licence permits delivery terms     pass
P16 eval sets decontaminated           pass
P17 tokenizer fertility measured       warn

# 2 blocking. run not started.
Where the build runs

Three places a build can run

The isolated network is what we built this for, and most of our engineering has gone there. It is not a prerequisite. Whichever of these three describes you, the same thing arrives at the end: safetensors and GGUF, the eval harness, and the configs to reproduce the run.

01 · inside the air gap

Nothing enters and nothing leaves

The case the rest of the market will not take.

Offline install from mirrored images, training on isolated hardware, delivery on physical media. No licence server, no telemetry, no callback anywhere in the path. This is the constraint the company was built around, and your firewall rules are the control, not our promises.

offline installread the security page →
02 · on your hardware

The data can be copied, but not offsite

Your GPUs, our engineering.

The run executes on the cluster in your data center or your VPC, against storage you control, with our engineers driving it. The right answer when residency rules bind the data but the network is not actually sealed.

your computesee the three rungs →
03 · on our GPUs

No compliance boundary at all

You just want the model.

Not everyone who needs a domain model is under supervisory rules. If nothing about your problem requires an air gap, we do not charge you for one: you bring the domain, we bring the compute and the training stack, and you still own the weights outright at the end.

from $15,000see the three rungs →
Bonacci Foundry · model portfolio

The models we build for ourselves

We pretrain from scratch, in house, on our own infrastructure. There is no checkpoint from another lab sitting underneath any of these. Every model ships with its evaluation harness and a written account of where it fails, because a model nobody can audit will not survive a deployment review in a regulated environment. The portfolio runs on the same pipeline a Foundry engagement uses, and it grows as we take on new domains.

Kautilyaa 9B finance

Grounded financial analysis. Reads live market data from context and quotes it exactly, or tells you it does not have the number. Pretrained on SEC EDGAR filings and financial news. 9B parameters with 2.6B active, Apache-2.0, commercial use permitted.

weights, evals and limits →
Open weights

CodeBERT‑Permissive code search

Search a codebase by describing what the code does, in plain English. Pretrained from scratch on 3.97B tokens of permissively licensed code, with every licence decision and every redaction written down. 110M parameters, runs on a CPU, Apache‑2.0.

weights, evals and limits →
Open weights

CyberAgent 9B internal tooling

Code and infrastructure generation for our own build pipeline. Pretrained in house on a public code corpus. It exists to run our tooling. The open code models are better at general code and we use those too, so this one stays internal.

Internal

What that looks like in use: an analyst drops a quarterly filing into the context window and asks for the segment margin. Kautilyaa returns the figure and the line it came from. Ask it for something the filing does not contain and it says it does not have that number. Across 150 test cases it produced no invented figures at all, which is the behaviour that decides whether a model is allowed anywhere near a client report.

From the build log

The useful output of a build is rarely the benchmark

On Kautilyaa we traced a silent optimizer bug in the training stack that had been quietly degrading every run: weight decay applied to normalisation gains, invisible in the loss curve, measurable only in the weights afterwards. Catching that class of failure before your compute is spent is what a Foundry engagement is actually for.

Commission a model Browse the weights
Your weights
delivered, not licensed
Zero
third-party AI calls in what we deliver
Air-gap capable
offline install, no callbacks
17
precheck gates before compute
Why thinkingdbx

Everyone else rents you a model. We hand you one.

The frontier labs will not train inside your network, and they will not hand over weights at any price. If your data cannot leave, that closes the category entirely. We build in the only place your data is allowed to be, and you keep what comes out. Teams with no such constraint get the same deal, they just have more options than you do.

your models

The weights end up yours

Delivered as safetensors and GGUF with the eval harness and the configs to reproduce the run. There is no per-seat AI tax and no renewal date, so nobody gets to decide next year what you are allowed to keep running.

audited builds

Checked before the compute burns

Corpus, configuration, and tokenizer checks run before a GPU hour is billed. When the run finishes we open the weights up and look at what is actually in them, then hand you the harness so you can repeat every check yourself.

data sovereignty

No outbound dependencies

Local inference by default, offline licence validation, your own SMTP and object storage. Nothing in the install path requires a call back to us. Block us at the perimeter and the software carries on working exactly as before.

no corpus required

Thin data still works

A few hundred real examples expanded into training volume, or a corpus we curate for you from public sources. Starting with no dataset changes what the build costs and how long it takes. Plenty of teams start there.

sized to serve

Small enough to actually run

We build in the 9B class, sparse where it helps, so the result serves on hardware you already own. There is no point delivering something that will not fit on the GPUs sitting in your rack.

honest audits

We will talk you out of it

Roughly half of our corpus audits conclude that retrieval or a fine-tune is enough, and we put that in writing. We are not charging for the week at present, so telling you to walk away costs us nothing.

Who this is for

Teams whose data legally cannot touch a public cloud

This is who we built for. If the answer to “can we send this corpus to a frontier lab?” is no, then buying a model is not available to you and building one is the only path left. Everything on this page assumes that constraint, which is why it also works for teams who do not have it and simply want a model of their own.

Defense & government

Classified and controlled-unclassified networks where air-gap is a hard requirement and offline delivery is the only install path.

Banking & financial services

Core banking, transaction, and PII data under supervisory rules that forbid third-party model processing.

Healthcare & life sciences

Patient records and trial data, where auditors care about where the data physically sits rather than how well it is encrypted in transit.

Critical infrastructure

Energy, utilities, and industrial operators running OT networks that are isolated from the internet by design.

For systems integrators and primes

We deliver as a subcontracted line item

If you have a pursuit with an on-premise or sovereign model requirement, we build it under your delivery. Your client ends up owning the weights outright. There is no telemetry and no callback in anything we hand over, so it should clear your security review without argument, and the audit report is written so it can go into your bid as it stands.

Send us the pursuit
The corpus audit · one week · free for now

It starts with one week that can end in us telling you not to build.

The audit reads your corpus, measures tokenizer fit, trains small models on your own data to see where the curve goes, and works out what a build would actually cost you in tokens and compute. Roughly half of them conclude that retrieval or a fine-tune is enough. You get the written report either way. We are running the audit free of charge for now, and you are committed to nothing at the end of it.

STEP 1

The audit week

Your domain, what data you actually have, where the build is allowed to run, and a scaling ladder on your corpus to see what size is justified.

STEP 2

Written report

The findings, a token budget, a recommended rung, and the eval targets a build would have to hit. Including the recommendation not to build, when that is the honest one.

STEP 3

The build

On our GPUs, on yours, or inside your air gap. The eval targets are agreed in writing before the run starts, and if the delivered weights miss them the remediation work is ours, not billed to you.

Separately available

The rest of what we make, none of it required

A Foundry engagement is complete when the weights are in your hands. Everything below is a separate product that exists because we needed it ourselves to build models. Three of them are open source, so if you are weighing up a regulated build with a small vendor, you can go and read the code today.

Bonacci Studio optional

If you also want the pipelines around the model: an agentic data platform that serves your weights on your own Ollama or vLLM and ships pipeline code to your Spark and Kafka. Self-hosted, air-gap capable, from $60,000 a year. Decline it and your model works exactly the same.

Platform

vibeCodeCLI free

Parallel agent fleets and persistent memory for coding and data engineering, in your terminal. Free forever, bring your own keys.

Live

ThinkingMemory

A memory database for agents. Intent goes in, the right context comes out, packed to a token budget.

Open source

ThinkingLanguage

A typed, compiled language where tables, streams, and models are native types.

Open source