Custom AI Models Financial Institutions Own and Run In-House
Ertas lets banks and financial institutions train custom models on their own examples, then run them on their own infrastructure or through the adaptive model endpoint, so the models behind your workflows are assets you own, test and version.
The Challenges You Face
Regulators Care Where Customer Data Goes
Banking supervisors, privacy law and outsourcing guidelines all put conditions on where customer data is processed. Sending every transaction narrative to a third-party AI API brings vendor risk reviews and data processing agreements for each new use case, which slows adoption of the most valuable work.
Model Risk Management Needs Models You Can Hold
SR 11-7 and similar frameworks expect institutions to document, validate and monitor the models used in decisions. A hosted model that can change without notice, and that you cannot download, is hard to validate and harder to keep validated.
Generic Models Miss Institution-Specific Categories
Every institution has its own transaction taxonomy, product names, risk tiers and house style. General-purpose models classify against a generic picture of finance, so their output needs heavy correction before it fits your categories.
Build vs. Buy Creates a Hard Trade-Off
Building an internal ML platform is a multi-year investment. Relying entirely on frontier APIs means paying per token indefinitely for work a smaller trained model could do. Institutions need a middle path.
How Ertas Solves This
Ertas Studio trains a custom model on examples of your own work: transaction narratives with the categories your analysts assign, filings with the fields your team extracts, customer queries with the answers your policy allows. Analysts can build, correct and approve those examples in Data Craft, Studio's guided dataset builder. Studio runs LoRA fine-tuning of open models such as Llama, Qwen, Mistral or Gemma on Ertas's cloud GPUs through a visual canvas, with several runs in parallel so you can compare configurations.
Training data is uploaded to Ertas for training, so prepare it to your own data-handling standards first and remove account identifiers the task does not need. Datasets are stored so runs can be repeated, and you can delete them.
The trained model is yours. Download it as GGUF or full safetensors and run it inside your own infrastructure with llama.cpp, Ollama or vLLM, where production data stays with you. Because you hold the model file and the dataset version it was trained on, your model risk team can test, document and re-validate it like any other model in your inventory. Meeting your regulators' expectations remains your institution's own work.
For a high-volume task already running on a frontier API, the adaptive model endpoint offers a gradual path: one API, a custom model that takes over a growing share of requests once it passes your evals, and hard cases left on the frontier model. It runs Ertas-managed, in your own cloud, or on-premise. Larger programs can start with an enterprise build: a free use-case audit, a three-to-four-week proof of concept, the build, then running and retraining.
Key Features for Financial Services & Banking
Analyst-Built Training Data
Data Craft lets analysts write and correct example exchanges, version them, and track which ones are approved for training, so the model learns your institution's categories from the people who apply them.
Versioned Datasets and Repeatable Runs
Data Craft versions your examples and Studio keeps the settings of each run, so a run can be repeated and each model traced back to the data and configuration that produced it.
Models You Hold
Every trained model is listed in Hub and downloads as GGUF for llama.cpp or Ollama, or as full safetensors for vLLM and other server runtimes. Run it behind your own firewall with no Ertas dependency at runtime.
Adaptive Model Endpoint
For one high-volume task, the adaptive model endpoint gives you a single API. A custom model trained on your own requests and your team's feedback takes over a growing share once it passes your evals, while hard cases stay on the frontier model you already use. Adopting it means pointing existing API calls at a new model name.
Why It Works
- On the QMSum meeting-summarization benchmark, a 1.2B model trained on Ertas scored 35.41 ROUGE-1 against 32.43 for the best of five hosted frontier models: a small model trained for one task can beat far larger general models on that task.
- Models export as GGUF or full safetensors, so they run on llama.cpp, Ollama, LM Studio or vLLM inside your own infrastructure.
- A self-hosted model changes only when you retrain it, so your validation team can test it on its own schedule.
Example Workflow
A mid-size bank wants to triage Suspicious Activity Report narratives so the highest-risk cases reach senior analysts first. The compliance technology team exports 3,000 historical narratives with the risk tier and typology each was assigned, removes account numbers and names the task does not need, and formats them as a JSONL dataset.
Experienced BSA analysts review the examples in Data Craft, correct inconsistent labels and approve the final set. The team trains three runs in parallel in Ertas Studio on different base models and compares them against a held-out set of recent narratives.
The chosen model is downloaded as a GGUF file and served on the bank's own GPU server behind the firewall. It pre-scores incoming narratives, analysts make every final decision, and the model risk team documents the model with the dataset version and run settings that produced it.
Related Resources
Ship AI that runs on your users' devices.
Free plan with 30 credits/mo, no card required. Paid plans from $10/mo USD.