01ingest
Source & contract
Pull from the systems of record with a documented contract: owner, refresh, retention, permitted use.
ModelsModels trained on
Models trained on
your distribution.
A foundation model has read the internet. It has never read your contracts, your tickets or your edge cases. We adapt it to your data and prove the difference with an evaluation harness you own.
02Hawthorn AI Labs
hawthorn@labs:~/train
$hawthorn train --base llama-3.3-70b --method lora
$shuffling 42,180 examples · 6 capability buckets
$epoch 3 · eval 71.2 → 89.8 (+18.6)
$safety suite 128/128 passed
$artifact → registry/models/finance-qa@v4
candidate promoted
- Typical lift
- +15–22 pts
- Iteration cycle
- Weekly
- Base options
- Open + frontier
- Serving
- vLLM / TGI
01 / What we build
Not every problem needs a 70B.
We choose the smallest model class that clears your bar, then make that model genuinely good at the job.
01
Domain assistants
Chat and reasoning models that know your products, policies, tone and constraints — without being told twice.
Explore02
Classifiers & routers
Small, fast models that decide what a request is, who owns it and which path it should take before any LLM is invoked.
Explore03
Extraction models
Structured output from messy inputs: contracts, invoices, clinical notes, tickets, logs — schema-constrained and validated.
Explore04
Retrieval & reranking
Embedders and cross-encoders tuned on your corpus so retrieval precision stops being the bottleneck for everything downstream.
Explore05
Document & vision
Layout-aware understanding for scans, forms, tables and screenshots where text-only pipelines quietly fail.
Explore06
Distilled specialists
A large teacher's behaviour compressed into a small, cheap student you can run at volume, on-prem, or offline.
Explore02 / How we train
Four methods. Chosen on evidence.
We run the cheapest experiment that can answer the question. Full fine-tuning is a last resort, not a default.
Supervised
Supervised fine-tuning
The workhorse. We build curated instruction data from your systems of record, filter it ruthlessly, and train until the held-out set moves.
- Curriculum-ordered examples, not a dump of tickets
- Deduplicated, decontaminated and split by entity, not by row
- LoRA for iteration speed; full-parameter for the final model
- Loss tracked per capability, not just as one aggregate
03 / Data
The model is the easy part.
Almost every failed fine-tune is a data problem wearing a training problem's clothes. This is where the weeks actually go.
02clean
Filter & deduplicate
Remove near-duplicates, PII you are not allowed to train on, and the 40% of rows that teach nothing.
03split
Entity-level splitting
Split by customer, document or account — never by row — so the eval set cannot leak into training.
04label
Expert review loop
Your domain experts adjudicate the hard cases; their decisions become both training signal and the rubric.
05train
Train & ablate
Small, cheap runs across data mixtures and hyperparameters before any expensive commitment.
06gate
Evaluate & promote
The candidate must clear every gate below. It either promotes or it goes back with a diagnosis.
04 / Evaluation
Every release has a gate.
A number nobody agreed on is not a metric. These are set with you before the first run and enforced automatically thereafter.
| Metric | What it measures | Release gate |
|---|---|---|
| Task accuracy | Does it do the job on held-out, entity-split data? | ≥ baseline + 5 pts |
| Groundedness | Are claims supported by retrieved or supplied context? | ≥ 95% |
| Schema validity | Does structured output parse and validate every time? | ≥ 99.5% |
| Refusal precision | Does it decline out-of-scope asks without blocking valid ones? | ≥ 98% |
| p95 latency | End-to-end response time under production load. | agreed per route |
| Cost / 1k runs | Fully-loaded inference cost including retrieval and tools. | budget ceiling |
| Safety suite | Injection, exfiltration, prompt-leak and jailbreak battery. | 0 criticals |
Serving & training stack
PyTorch/TRL/vLLM/Hugging Face/MLflow/Weights & Biases/Ray/DeepSpeed/ONNX/TensorRT/Airflow/Postgres/
Custom modelsBring a baseline.
Bring a baseline.
We'll beat it.
Send us a held-out sample and the metric you care about. You get a scored baseline before any contract is signed.
Typical first response within 24 hours · NDA-friendly