Tell our agents what you need. They build your specialized model.
Describe the job in plain language and answer three or four questions. Our agents pick the base model, prepare the data, train, test on your real data and run every job on the fewest GPUs. You get a message when your model is ready, and the model is yours.
- Task
- Text classification, 12 categories + 3 extracted fields
- Target
- F1 ≥ 0.90 on your real data
- Data
- Your examples + filtered synthetic data
- Delivery
- Export into your systems
A model that knows your work usually takes experts, data and GPUs you don't have
Too many experts
Choosing a base model, preparing data, training and testing each need specialists. Our agents do that work for you.
Too little data
Labelled examples are scarce, sensitive and slow to create. Our agents use what you have and generate the rest.
Too much compute
Fine-tuning usually means GPU clusters, and regulated data has to stay in approved environments. We run lean, and in-Kingdom.
You describe the job. Our agents build the model. You own it.
One conversation. No menus, forms or wizards, and nothing for your team to do after the brief.
Describe the job
Say what you need, loosely or precisely, in the console or from your own agent through the API. Answer three or four questions, then get on with your day.
The agents build it
They pick the base model, prepare the data, train, and test on your real data. Follow progress and cost any time.
Receive a tested model
When it meets your target you're notified in-app, by email, WhatsApp, SMS or webhook. It's delivered into your systems or hosted for you.
Product preview. Figures shown are an example.
Hand it over completely, or specify every detail
You never pick a mode. The platform reads how much control you want from what you say.
Detect garbage in areas.
We decide everything, ask three or four questions, and get going.
Small enough to run on a phone, Arabic first.
You set preferences and we optimise within them.
Fine-tune Qwen-7B with LoRA, F1 at least 0.9.
We build exactly that, and still test it against your target.
We only suggest a change when it's worth it. If you name a model, we build it and test our own best candidate alongside. We'll propose the alternative only if it's at least 25% cheaper at the same accuracy, or at least two accuracy points better at the same cost, and we'll show you the results. Say "exactly" and we won't second-guess.
See the plan before any spend
A plan card shows what will be built, the target, and the estimated cost and time. You can edit or stop it, or require approval first.
Change your mind mid-build
Send a message like "also handle Urdu" and the plan updates, showing you what changed.
Your second model is faster to brief
We remember your past briefs, data sources and targets. "Same as the ticket model, for complaints" is enough.
Specialist agents, one job each
The work of an ML team, done by agents that never stop for the weekend.
Concierge
Talks with you, asks only what it can't work out, writes the plan and keeps you updated.
Model Scout
Searches Hugging Face and other open sources for base models that fit your job, with size, licence and cost.
Data agent
Uses your data, generates synthetic data, or both. Filters for quality, removes duplicates and handles personal data.
Training agent
Fine-tunes the chosen model efficiently, using proven methods such as 4-bit QLoRA.
Evaluation agent
Scores every candidate on your real data for accuracy, speed and cost. Nothing ships below your target.
Delivery agent
Packages the model, deploys it into your systems or hosts it, then monitors and retrains it when you ask.
GPU Optimizer
Sizes every job to the smallest GPU that fits and tracks the cost of each one.
Notifier
Sends "your model is ready", "more examples needed" or "target not met" on your channel, with one-click next steps.
Specialized models that run on far fewer GPUs
Our GPU Optimizer applies techniques proven in published research to every build, so you get the savings without needing the expertise.
Any industry, any high-volume task
We start with financial services and telecoms, where large volumes of Arabic and English text meet strict data rules.
Transaction exceptions
Classify exceptions into categories and extract the key fields, in Arabic and English.
Trouble tickets
Classify and route customer and network tickets automatically.
Complaints
Sort and prioritise complaints so they reach the right team quickly.
KYC documents
Extract and check information from know-your-customer documents.
HR requests
Classify and answer routine HR requests and records.
Litter detection
Spot garbage in images of public areas. Coming with vision support.
Text first, with more data types on the way
Text
Classification, extraction and embeddings, in Arabic and English.
Vision
Image classification and object detection.
Audio & speech
Models that work with recordings and calls.
Documents & charts
Forms, reports, charts and graphs.
Tabular & time series
Structured records and signals over time.
Your agents can order specialized models from ours
We build models, not agents, so we never compete with you. Embed Modelsmith once and order tested models for each of your customers.
- REST API. Send a brief and a few examples, alone or with structured preferences. Get a tested model back.
- MCP tool server. Any agent framework that speaks MCP can order a model as a tool call.
- Webhooks. Build updates and the "model is ready" event go straight into your systems.
# your agent sends a brief POST /models { "brief": "Classify complaints by product and urgency", "languages": ["ar", "en"], "target": { "f1": 0.9 }, "examples": "complaints_sample.jsonl" } # ...later, on your webhook { "event": "model.ready", "f1": 0.92, "endpoint": "https://…" }
Sketch for illustration. API reference available at launch.
Your data stays home. Nothing ships until it's accurate.
Run on our in-Kingdom cloud in your own isolated tenant, or run the whole agent team inside your own environment with the Enterprise edition.
Quality
- Every model is tested on your real data. Delivery is blocked below your target.
- Synthetic data is always quality-filtered, and real data is always used for testing.
- If a target can't be met, you're told what was achieved, why, and your options.
Security
- Agents act only inside your tenant, using approved tools.
- Encryption at rest and in transit, role-based access and a secrets vault.
- Every agent action is logged and exportable.
- Your data never trains or informs another customer's models.
Built for regulated teams
- Keeping data and models in-Kingdom supports your obligations under PDPL, SAMA and CST rules.
- Approval steps before delivery, retraining or export are available.
- The models you receive are yours to run wherever you choose.
Pricing built around the models you need
A platform subscription plus a monthly fee for each specialized model, covering hosting or support, monitoring and retraining. Talk to us for a quote.
Self-serve
For fintechs and growing companies. Sign up and brief the agents yourself, starting with one model.
Contact usSeveral models, your environment
For banks, insurers, telecoms and government. Includes the Enterprise edition to run inside your own environment.
Contact usProve it on your data
For larger organisations that want to see results on their own data before committing.
Contact usFrequently asked
When can I start?
We're launching soon. Join the waitlist and we'll invite you as soon as your access is ready, in waitlist order.
Are you building agents that do my work?
No. Our agents build models. You get a specialized model you own; what it does in your business is up to you.
How much work is it for my people?
A brief and three or four answers. The agents run the whole build and tell you when your model is ready.
Do I have to give you my data?
No. The agents can work from a description, with synthetic data filling the gap. Real examples improve the result and are always used for testing.
Will it spend my budget without asking?
You see a plan with the estimated cost and time before any GPU is used, and you can edit or stop it. You can also require approval first.
Where does the model run afterwards?
Wherever you want: exported into your systems or hosted by us. The Enterprise edition keeps everything inside your environment.
What happens when a new AI model comes out?
Nothing breaks. Every new open model becomes another option our Model Scout can choose for you.
Is it only text?
Text comes first. Vision follows, then audio, documents and charts, and tabular data.
Be first to brief the agents
We're opening Modelsmith to a limited group first. Tell us about your first model and we'll reach out before launch.
- Priority access when we open, in waitlist order.
- Your first brief reviewed by our team, so it's ready to build on day one.
- Talk to us first about plans, pilots or the Enterprise edition.