Sinewgrid Finance trains financial language models and quant research code generators.
Two related jobs: adapting a general language model to finance with continued pretraining and fine-tuning, and assistants that write and revise research code against backtests. This domain is at the exploring stage, and nothing here is investment advice.
Two jobs in one domain.
Sinewgrid Finance covers two related jobs. The first is adapting a general language model to finance, with continued pretraining and fine-tuning on licensed market data, microstructure data and financial text, so that it handles filings, transcripts and research material with the right vocabulary and context. The second is an assistant that writes and revises quantitative research code against backtests: it drafts a strategy, reads the result, and revises, which means a single idea can trigger many model calls. Both are compute-hungry in different ways. Adaptation needs long multi-node runs, while the code-and-backtest loop and event-driven inference, which spikes around earnings, central-bank decisions and market shocks, need burst capacity that is planned in advance, with firm deadlines.
Finance adds constraints that shape how we work. Data comes from licenses the customer holds or we obtain, and the terms are recorded and respected. A customer’s own research code and results stay in its tenant, under its keys. Models are judged on point-in-time, out-of-sample tests with leakage controls, and a customer’s model-risk review comes before anything touches a live process. We would report out-of-sample accuracy, cost per answer, counting every token including verification and outside model calls, and latency during a simulated burst. We do not publish returns or trading results, the model does not trade on our side, and nothing on this site is investment advice. This domain is at the exploring stage.
Where the capacity goes.
Adapting a large model
Continued pretraining and fine-tuning on market data, microstructure data and financial text needs long multi-node runs.
The code-and-backtest loop
An assistant that writes a strategy, reads the backtest and revises it makes many model calls for every idea.
Bursts around events
Earnings, central-bank decisions and market shocks send inference demand up sharply, with firm deadlines.
Checking adds tokens
Verification and self-correction steps mean each answer costs more hidden tokens than a plain reply.
What the models learn from.
Licensed market data
Prices, order-book and reference data under license, with the license terms recorded.
Financial text
Filings, transcripts and research text the customer is entitled to use.
Customer data
A customer’s own research code and results, kept in its tenant under its keys.
From a question to a checked result.
- Define the taskA question type or a research workflow, and the test that will judge it.
- Adapt the modelContinued pretraining or fine-tuning on the licensed data.
- Evaluate out of samplePoint-in-time backtests and out-of-sample checks, with leakage controls.
- Review and deployModel-risk review before anything touches a live process.
Steady training, bursty and urgent inference
Reserved capacity for training. Burst capacity, planned in advance, for event-driven inference.
See the architectureNo performance claims
We do not publish returns or trading results. Models are evaluated on point-in-time, out-of-sample tests, and customer model-risk review comes before any live use.
Read the commitmentsA first engagement, scoped small.
First engagement
One research workflow, for example filing analysis or factor code generation, with a clear out-of-sample test.
What the customer brings
Licensed data, the workflow, and a model-risk contact who signs off before any live use.
What we bring
Model adaptation, the evaluation harness, and a capacity plan that covers event bursts.
Measures we commit to publishing.
Each result is reported against these measures, which are defined before the first run.
Out-of-sample accuracy
Task accuracy on point-in-time data the model never saw, with leakage controls described.
Cost per answer
Total tokens, including verification steps and any outside model calls, for each completed task.
Latency at peak
Response time during a simulated burst, against the customer deadline.
Plain answers.
Do you give investment advice or returns?
No. We do not publish returns or trading results, and nothing on this site is investment advice.
Where does the data come from?
From licenses the customer holds or we obtain. License terms are recorded and respected.
Can the model trade?
Not on our side. Any link to a live process is the customer's decision, after their own model-risk review.
Financial use brings data-license, model-risk and regulatory questions. We treat them as design inputs from the start. Related: the Platform: the control plane handles burst capacity and failover. Status: Exploring
Planning a training run?
Tell us the model, the data and the schedule. We reply with a capacity plan and the evidence behind it.