We built Baseweight because every AI shop has one tool—and recommends it for everything.
Every AI services company has a specialty. Fine-tuning shops recommend fine-tuning. RAG consultancies recommend RAG. Prompt engineers recommend more prompting.
This isn't because these are bad companies. It's because when you have one tool, every problem looks like it needs that tool.
The result: teams burn $10–50k and months discovering the wrong technique was applied. Then they start over—if they can justify the investment internally.
Baseweight exists to fix the sequence. Diagnose the problem across the full adaptation stack. Then—and only then—select and execute the right technique.
Our team has built domain adaptations across retrieval, fine-tuning (LoRA, QLoRA, full), instruction tuning, preference optimization (DPO), and model compression. Our value is in the diagnosis, not the technique—so we never default to a single approach.
Every engagement produces artifacts you own: weights, training recipes, eval sets, deployment configs. We hand you the methodology so your team can maintain and extend it. If you want us to stick around, retainers are available—but dependency is never the business model.
What We Believe
Diagnosis before prescription
The technique should follow from the problem, not the other way around.
Ownership over dependency
You should own your model, your weights, and your methodology. Full stop.
Patterns over bespoke
Every engagement contributes to a growing library of reusable templates. You benefit from patterns already tested in other domains—not a blank-slate build.
Honesty over revenue
If your problem is solvable with better prompting, we'll tell you. We'd rather earn trust than bill hours.
Founded by Philip Stevens
12 years in applied machine learning, including production ML at Agoda and Quantcast. Independent consultant since 2023, focused on post-training, eval design, RAG hardening, and parameter-efficient fine-tuning (LoRA/QLoRA).
Baseweight was built from a pattern that kept repeating in consulting engagements: teams choosing adaptation techniques before diagnosing the problem. The diagnostic-first model exists because that sequence works—and the reverse doesn't.
Independent consultant since .
- Post-training
- Eval design
- RAG hardening
- LoRA / QLoRA
- DPO alignment
- Model compression
Not ready to book a call?
Get the free AI Adaptation Readiness Checklist and Technique Selection Guide. Figure out where you stand before we talk.
No spam. Unsubscribe anytime.
Ready to diagnose your adaptation stack?
30 minutes. No cost, no commitment. We'll tell you which technique your domain data actually needs.
Book an Intro Call