AI Model Development
General models are trained to be decent for everyone and perfect for nobody. When your work depends on words, labels or decisions a general model gets wrong, a model trained for the job does better, and often runs smaller, cheaper and on your own hardware. I train those models, with speech recognition at the front, and advise teams who train their own.
Flagship: ASR fine-tuning
Speech recognition is where I do most of my own training work, and where a fine-tune makes the most visible difference.
ASR Fine-Tuning
Speech recognition models fine-tuned for your vocabulary, accents and speakers who switch languages mid-sentence, then deployed offline, on a public API or as a private internal API.
See ASR fine-tuning →- Niche vocabulary: legal, medical, technical, brand and project names
- Under-represented accents
- Code-switching between languages
- Base model chosen by workload and target device
- Real or synthetic speech training data
- Retraining as your vocabulary changes
My own fine-tunes are public: see the open-source ASR models catalogue.
Fine-tuning beyond speech
The same process applies to text: start from an open model, train it on your data for one job, and measure it against what you use today.
LLM Fine-Tuning
Open-weight LLMs such as Llama, Qwen, Mistral and Gemma fine-tuned with LoRA for your domain, output format and tool calls, or distilled from a frontier model into one you host.
LLM fine-tuning → CC-SVC-15Classifier Training
Custom text classifiers for routing, triage, intent, document type and moderation, trained on your labels and cheap enough to run on every record.
Classifier training → CC-SVC-10Small & Specialised Models
Embedding models and rerankers tuned for your retrieval, extraction models, and small language models for on-premises and on-device use.
Small & specialised models →Data and evaluation
Behind every model I train: the dataset it learns from and the evals that show it beats what it replaces.
Datasets & Training Recipes
Dataset preparation and blending, drafting and reviewing training plans, and version control and release cadence for models you deploy internally.
Datasets & training recipes → CC-SVC-12Benchmarks & Evals
Evals for your own tasks, a stable frame of comparison for new model releases, and evaluation run as an independent service.
Benchmarks & evals →Working with me on a trained model
Whichever model it is (speech, language, classifier or a small specialised model), the engagement has the same shape: a measured baseline, a trained model you own, and the evidence that it does better.
Who it's for
- Teams whose work depends on words, labels or decisions a general model gets wrong
- Pipelines paying per call for a narrow task a small model could do
- Organisations that need models offline or on their own infrastructure
- Teams training in-house who want their data and recipe checked
What you get
- A trained or fine-tuned model, weights yours
- Before-and-after evals on your data
- The training data and recipe, for retraining
- Deployment into your pipeline: offline, internal API or hosted
- A retraining and release plan
What I need from you
- Examples of the task, and what currently goes wrong
- Whatever data you already have
- Where the model has to run
- One person who can judge outputs
How it runs
Typically 3–8 weeks depending on the model and the data.
- Weeks 1–2
Scope and baseline
Define the task and measure the current model on your data.
- Weeks 1–3
Data and plan
Training data plan, base model and method.
- Weeks 3–6
Train and evaluate
Train and compare against the baseline on held-out data.
- Weeks 5–8
Deploy
Ship it into the pipeline and hand over.
Training work has no public price list. Each model is scoped after a short call and quoted as a fixed price, paid against milestones.