Post-train a frontier model on your domain.
When the problem is language and judgment, we take a strong base model and post-train it on your data, voice, and format — supervised fine-tuning and preference optimization — and prove the gain on held-out evals before it ships.
How a base model becomes yours.
A disciplined sequence, with an eval gate before anything ships.
Base model
Start from the right frontier or open model for the task, licensing, and deployment target — chosen, not defaulted.
Supervised fine-tuning
Teach the model your task, format, and tone from curated input-output pairs — the fastest path to a specialized model.
Preference optimization
Align behavior to human judgment with RLHF or DPO — rewarding the answers your experts prefer, penalizing the ones they don't.
Evaluation
Prove the tuned model beats the base on a held-out benchmark for quality, safety, and regression before it goes near production.
Distillation & serving
Compress into a smaller, cheaper model where you can, then serve behind an API with routing and monitoring.
The methods we reach for.
Supervised fine-tuning (SFT)
Curated demonstrations teach task, format, and domain language — the highest-leverage first step for most specialization.
RLHF
Reinforcement learning from human feedback trains a reward model on preferences, then optimizes the policy against it.
DPO
Direct preference optimization skips the separate reward model — simpler, stable, and often enough.
LoRA / PEFT
Parameter-efficient tuning adapts a large model by training small adapters — cheaper, faster, and easy to swap.
Distillation
Transfer a big model's behavior into a small one to cut latency and cost while keeping most of the quality.
Continued pretraining
When your domain language is genuinely unusual, keep pretraining on your corpus before task tuning.
Prompt, retrieve, or fine-tune?
Fine-tuning is powerful and often unnecessary. Here is how we decide.
| Approach | Best for | Cost to change | Watch out for |
|---|---|---|---|
| Prompt engineering | Behavior a strong model can already do with the right instructions and examples. | Minutes | Gets fragile as prompts grow; token cost on every call. |
| Retrieval (RAG) | Knowledge that changes or must be cited — facts live outside the model. | Hours | Retrieval quality is the ceiling; won't change style or skill. |
| Fine-tuning | A consistent skill, format, or tone the base model doesn't have — proven at scale. | Days to weeks | Needs quality data and re-training as the world moves. |
For most knowledge problems, retrieval is cheaper, more current, and easier to audit than training. We fine-tune when a task needs a skill or consistency the base model lacks — and we tell you when it doesn't, instead of selling a training project you don't need.
Bring us a hypothesis. Leave with a system.
Tell us what's eating your team's time. We'll give you an honest read on whether AI is the right tool — and if it is, a scoped v1 with a timeline and cost.