GRAYBOOK LABS

Capabilities

Custom LLMs, RAG and NLP, built end to end.

From the first look at your data to a model serving answers in production, one team designs, trains, evaluates and runs it.

01 · Custom LLMs

Language models trained and tuned for one job.

General chatbots don’t know your instruments, your rules or your vocabulary. We pre-train, fine-tune and align models on the data that matters to you, then shrink them until they run fast on hardware you control.

  • Continued pre-training on market, filing and research text
  • Supervised fine-tuning, LoRA and QLoRA adapters
  • Preference tuning (DPO) so answers sound like your desk
  • Distillation and quantization for low-latency serving
  • Evaluation suites built from your own questions

02 · RAG architectures

Answers that cite their sources.

Retrieval-augmented generation lets a model answer from your documents instead of its memory. We design the whole pipeline, from parsing and chunking to hybrid search, reranking and citations, so every answer points back to where it came from.

  • Document parsing for PDFs, tables, filings and transcripts
  • Embeddings and vector indexes tuned to your corpus
  • Hybrid keyword + vector search with cross-encoder reranking
  • Live data joins: prices, positions and trade history
  • Citations on every answer, and refusals when the data isn’t there

03 · Natural language processing

Structure from unstructured text.

News, filings, central-bank statements and chat logs are full of signal and noise. We build NLP pipelines that read them at machine speed: pulling out entities and numbers, tagging events and sentiment, and handing clean data to the systems that trade on it.

  • Named-entity recognition for tickers, companies and people
  • Event and sentiment classification on news and filings
  • Summaries of earnings calls and policy statements
  • Multilingual pipelines with domain vocabularies
  • Streaming inference that keeps up with the feed

The tools we know best.

Deep experience where it counts: training, scaling and serving models, not just prompting them.

PyTorch

Every model we build is trained, fine-tuned and evaluated in PyTorch, from LoRA adapters on a single GPU to full fine-tunes across a cluster.

DeepSeek

We fine-tune and serve open-weight DeepSeek models, including mixture-of-experts reasoning models, on infrastructure you control.

Megatron-LM

For the largest runs we use NVIDIA’s Megatron-LM, with tensor, pipeline and data parallelism, to train models too big for any one GPU.

The hardware, clouds and platforms we build on

How we work with you.

Short loops, measurable results and no black boxes. You see answers on your own data early.

  1. 1

    Discovery

    We start with your questions, your data and what a good answer looks like, then agree on how we’ll measure it.

  2. 2

    Prototype

    A working pipeline on a slice of your data in [2–4] weeks, so you can see answers before committing.

  3. 3

    Train & evaluate

    Fine-tuning, retrieval tuning and red-teaming against your evaluation set until the numbers hold.

  4. 4

    Deploy & monitor

    Served on your hardware or ours, with logging, drift checks and a team that stays on it.

Questions we get asked.

Something else? Ask the team.

Does our data leave our environment?

It doesn’t have to. We can train and serve entirely on your hardware or in your cloud account, and nothing is used to train models for anyone else.

Do you build on open-weight or closed models?

Both, depending on the job. Open-weight models such as DeepSeek give you full control and can run on your own GPUs; we use hosted models where they are clearly better and the data allows it.

How do you stop a model from making things up?

Retrieval with citations, calculations done in code rather than by the model, evaluation sets built from real questions, and answers that say “I don’t know” when the sources don’t support one.

Is Vig financial advice?

No. Vig explains, calculates and helps you reason about your own trading. Decisions, and the risk, stay with you.

Can we use Vig outside the Graybook platform?

[Vig is built into the Graybook platform today. Talk to us about access for your firm.]

Bring us your hardest question.

We’ll show you what a model trained on your data can do with it.

Work with us →