Fine-tuning that has to beat its base to ship.
LoRA and preference tuning on open-weight models, judged blind against the untuned base on prompts that were never in training. The eval report is part of the deliverable, and so is the leak test.
Judgment, voice and honesty.
A tuned model should bring the interpretation while your systems bring the live facts. That split is what makes it safe to run unattended.
Domain judgment
Frameworks for one job, from B2B positioning to CFO-office advisory, learned from expert exemplars rather than the open web.
Brand voice at scale
Writers set the narrative; the model holds the voice across thousands of assets without drifting into generic copy.
Grounded arithmetic and calibrated refusal
Every derived figure traces to the numbers in the prompt. Time-sensitive facts get a decision framework and a pull-list, not a guess.
Fabrication resistance
Invented statistics and unsupported claims are a failure mode we test for explicitly, with traps in every evaluation set.
The right technique for the task.
Supervised fine-tuning
LoRA and QLoRA adapters on open base models. Small files, Apache-2.0 when public, portable to any serving stack.
Preference tuning
Align the model to what your experts prefer: terminology, tone, what an elite answer looks like versus a merely correct one.
Reinforcement fine-tuning
For tasks with a verifiable answer, such as NL2SQL, structured extraction or rule-following, reward the outcome directly.
Evals as a product
Blind pairwise judging by independent judges, fabrication traps, calibration probes, leak tests. Reproducible from the published sets.
Two adapters, public weights, public evals.
B2B Marketing v2
The copywriting brain behind our marketing AI employees. Positioning briefs in, on-brand B2B assets out.
- LoRA adapter on Qwen3-32B · Apache-2.0
- Blind-judged 7.16 vs 4.79 for its base model
- Preferred head-to-head over base in 86% of matchups
- Zero-leak verified
CFO Brain v1
Advisory brain for Indian mid-market finance. Your agent brings the live numbers; the model brings the judgment.
- LoRA adapter on Qwen3-32B · Apache-2.0
- Blind-judged 5.50 vs 3.79 for its base model
- Preferred head-to-head over base in 75% of scenarios
- Calibration probes 8/10 vs 5/10 for base
Common questions.
Do we have to hand over our data?
Training can run inside your VPC on the Own tier. Either way, client and brand data is scrubbed before training and verified absent after it, and nothing of yours goes into public weights.
Which base models do you use?
Open-weight families such as Qwen, Gemma, Llama and DeepSeek, chosen for the job and the hardware. We pick the base on our own evals, not the leaderboard.
Who owns the result?
Private adapters are yours. Public releases ship under Apache-2.0 with their evaluation sets, like the two above.
How long does it take?
Weeks, not quarters. Most of the time goes into the dataset and the evals; the training run itself is the short part.
Where does it run afterwards?
Wherever you choose on the Compute page: owned GPUs, cloud GPUs or your own keys, monitored on the cockpit.
Bring a task the API model gets wrong.
We will show you the dataset, the tuning plan and the eval it has to pass. No pitch deck. Just a technical conversation.