Daniel Rosehill · AI Model Training & Agentic AI · Jerusalem
I train the models
your pipelines run on.
Most AI consultants connect you to someone else's model. I train small models for your job, with speech recognition (ASR) fine-tuning as my speciality, and build the agents and workflows they run in.
Workflow Automation
Workflows built around your process that move work between your systems, with an AI step only where judgement is needed.
Learn more →Agentic AI
Agents that act instead of just answering: they read your context, use your tools and finish the job, with people in the loop where it matters.
Learn more →MCP
MCP servers give an agent typed, permissioned access to your CRM, tickets, files and databases, so it can take action safely.
Learn more →Fine-Tuning
Small models trained for one job on your own data. Speech recognition is my speciality: on code-switched speech, wrong words fell from 13.7% to 4.9%.
Learn more →Datasets
Your recordings, documents and corrections turned into training and test sets that you own and can reuse.
Learn more →Benchmarks & Evals
The model you use, or could use, tested on your own data, so you know how it performs before you build on it.
Learn more →- 13.7% → 4.9% wrong words on code-switched speech
- You own the model, the data and the code
What changes for you
AI that fits your work, not the other way round
Transcripts that get your words right
General speech models mishear product names, jargon and mixed-language speech. On my own fine-tuned models, wrong words on code-switched speech fell from 13.7% to 4.9%.
ASR fine-tuning →Your data stays where you put it
A small model trained for one job can run offline, on your own servers or behind your own API, so recordings and documents never have to leave your network.
Deployment options →AI that finishes the job
Agents and workflows that file the ticket, update the record and send the summary, instead of a chat window someone has to copy answers out of.
Agents & MCP →Something you own outright
You keep the model weights, the training data and the code. A vendor price change or model update doesn't change how your pipeline behaves.
How I work →How it fits together
One system, trained on your work
Trained models do the fast, exact steps; an agent with tools does the rest. Your own data keeps the models improving.
Your side
Inputs
Work arrives as it does today: recorded calls, voice notes, an inbox, documents.
Automation & Workflows
Workflow
Workflow automation picks each item up, moves it between systems and calls a model only where judgement is needed.
Explore workflow →AI Model Development
Speech model
A speech model fine-tuned on your vocabulary, accents and code-switching turns audio into text it gets right.
Explore speech model →AI Model Development
Classifier
A small classifier trained on your categories labels each item in milliseconds. Confident calls go straight to your systems without an LLM; the rest go to the agent.
Explore classifier →Agentic AI
AI agent
An agent handles the items that need reasoning: reading context, deciding what to do, drafting the reply.
Explore ai agent →Agentic AI
MCP tools
MCP servers give the agent typed, permissioned tools for your systems, so it acts instead of just answering.
Explore mcp tools →Your side
Your systems
The work lands where it belongs: a ticket filed, a record updated, a summary sent. Every run and every correction is kept as future training data.
Your side
Your data
Every run leaves logs, and every human correction is a labelled example. That is the raw material for training.
AI Model Development
Datasets
I turn your recordings, documents and corrections into training and test sets you own.
Explore datasets →AI Model Development
Fine-tuning
Models are trained or fine-tuned for one job each, small enough to run on your own hardware if you need to.
Explore fine-tuning →AI Model Development
Evals
Each model is benchmarked against the one it replaces on your own test set. Only a measured improvement ships back into the pipeline.
Explore evals →In production
-
Inputs Your side
Work arrives as it does today: recorded calls, voice notes, an inbox, documents.
-
Workflow Automation & Workflows
Workflow automation picks each item up, moves it between systems and calls a model only where judgement is needed.
Explore workflow → -
Speech model AI Model Development
A speech model fine-tuned on your vocabulary, accents and code-switching turns audio into text it gets right.
Explore speech model → -
Classifier AI Model Development
A small classifier trained on your categories labels each item in milliseconds. Confident calls go straight to your systems without an LLM; the rest go to the agent.
Explore classifier → -
AI agent Agentic AI
An agent handles the items that need reasoning: reading context, deciding what to do, drafting the reply.
Explore ai agent → -
MCP tools Agentic AI
MCP servers give the agent typed, permissioned tools for your systems, so it acts instead of just answering.
Explore mcp tools → -
Your systems Your side
The work lands where it belongs: a ticket filed, a record updated, a summary sent. Every run and every correction is kept as future training data.
Training loop
-
Your data Your side
Every run leaves logs, and every human correction is a labelled example. That is the raw material for training.
-
Datasets AI Model Development
I turn your recordings, documents and corrections into training and test sets you own.
Explore datasets → -
Fine-tuning AI Model Development
Models are trained or fine-tuned for one job each, small enough to run on your own hardware if you need to.
Explore fine-tuning → -
Evals AI Model Development
Each model is benchmarked against the one it replaces on your own test set. Only a measured improvement ships back into the pipeline.
Explore evals →
Improved models go back into the pipeline.
Hover or tap a step to see what it does.Tap a step to see what it does.
Capabilities
What I Do
Agentic AI
AI that does things: agents with access to your knowledge and tools, connected to your systems through MCP and put in front of your team through chat.
Explore Agentic AI →AI Model Development
Models trained or fine-tuned for one job, when a general model keeps getting the words, labels or decisions wrong. Speech recognition is where I specialise.
- ASR Fine-Tuning
- LLM Fine-Tuning
- Classifier Training
- Small & Specialised Models
- Datasets & Training Recipes
- Benchmarks & Evals
Automation & Workflows
Workflows and pipelines built for your process, with AI steps only where judgement is needed, and the reporting that shows they work.
Explore Automation & Workflows →Across all three: Consulting & PlanningUpliftDeployment OptionsMaintenance & Support
Clients I’ve worked with
Insights
Notes from the work
Write-ups on fine-tuning speech and language models, building agents and MCP servers, and the architecture around them.
Have a task a general model keeps getting wrong?
Send me the requirements: what the job is, the data you have and where it has to run. I'll tell you whether a trained model, an agent or plain automation is the right fix, and what it would take.