lilypad
Sign inGet started

Your data. Your model.Your weights.

Turn your data into a custom AI model. Lilypad handles everything from training to deployment.

From your data to your own AI.

Four steps, all in one conversation. Everything shown here comes from a real run on BANKING77, a public customer-support benchmark.

01Connect

Connect your data. Lilypad prepares and structures every training example.

Connect Gmail, HubSpot, upload files, or ask Lilypad to help build a dataset.

BANKING77, customer supportReady for training
All 10,003Included 9,999Excluded 4
  • I can't seem to be able to use my cardcard_not_workingDuplicate
  • I put the wrong pin too many times and now…pin_blocked
  • Where can I withdraw money from?atm_support
  • Do I need to go to a physical bank to change…change_pin
Excluded. Exact copy of an earlier example. The first copy is kept.
02Train

Describe the behavior you want. No ML expertise required.

Lilypad chooses the model, prepares the training, and evaluates the results.

BANKING77 Classifier, v1Completed
  1. PreparingWrites 8,993 training examples and uploads them
  2. Waiting for a GPUYour run waits in line until a GPU is free
  3. Training3 passes over your examples
  4. Saving the modelThe trained weights are compressed and stored
  5. ReadyYou test it on a test endpoint
Finished. The model is ready to test.Took 20 min
03Test

Tested on data it never saw, before anything ships.

Lilypad scores your model against the base model on examples held out from training.

Official test, 3,080 casesCompleted
BANKING77 Classifier, v1Fine-tuned92%
GPT-6 LunaOpenAI model82%
Qwen3.5 9BBefore training72%
+20 ptsover the same model before training
Reliable: the gap is well above the margin of error
04Deploy

Deploy it, or download your weights.

Use your model through an API at inference cost, with zero Lilypad markup, or export your LoRA weights.

PlaygroundReady to test
Customer messageI can't seem to be able to use my cardRun
Your modelcard_not_working128 ms
Download LoRA weightsadapter.safetensors

A model for the way you work.

Train on your best sales emails to generate personalized outreach in your team’s writing style.

Illustrative example

Trained on 1,936 emails that got a reply

Outbound WriterCustom-trained
InputProspect: Dana Kim, VP Operations at Fieldline Logistics. Signal: opened 3 dispatcher roles this month.
OutputHi Dana, three dispatcher roles in a month usually means the routing board is doing the planning. Worth 15 minutes to compare notes? Sarah

Learn from resolved support tickets to classify requests and draft better responses.

Illustrative example

Trained on 8,412 resolved tickets

Support SpecialistCustom-trained
InputI was charged twice for invoice INV-20931 and my card went over its limit. Can you refund it today?
OutputCategory: Billing · Priority: High Hi Maya, sorry about the double charge. I’ve flagged the duplicate for a refund today; it usually shows in 2 to 3 business days.

Turn approved marketing content into an AI that consistently writes in your brand’s voice.

Illustrative example

Trained on 640 approved posts

Brand VoiceCustom-trained
InputAnnounce that dark mode is live in the mobile app. Two sentences, for Instagram.
OutputYour late-night planning just got easier on the eyes. Dark mode is live; update the app and flip the switch in Settings.

Train on validated examples to extract structured information from documents, conversations, and workflows.

Illustrative example

Trained on 2,200 validated documents

Data ExtractorCustom-trained
InputNorthbay Clinics · Invoice 2291 Services rendered in October. Total due: $4,180.00, payable by Nov 15, 2026.
JSON output{ "vendor": "Northbay Clinics", "invoice": "2291", "total": 4180.00, "due": "2026-11-15" }

Your personal ML engineer.

From messy data to a production-ready model. Just ask.

What you did

What Lilypad did

  1. I want a model that writes emails like my best sales rep.

  2. Imported 4,812 email threadsGmail · sarah@northwind.io
  3. Removed 837 duplicates and signatures
  4. Masked 38 phone numbers

  5. Excluded 1 example

  6. Recommended Llama 3.1 8BIt writes short, personal emails well and is inexpensive to fine-tune.

  7. Approved the price, $18.40

  8. Trained for 42 min
  9. Tested on 60 prospects it hadn't seen43 of 60 test emails were judged closer to Sarah's real writing than the base model's.

Judged by comparing each draft with the email Sarah actually sent. Illustrative demo data.

Built for you. Owned by you.

  • Dataset preparationLilypad
  • Model selectionLilypad
  • Fine-tuningLilypad
  • EvaluationsLilypad

No ML expertise required

Lilypad handles dataset preparation, model selection, fine-tuning, and evaluations.

Input tokens
GPU cost
Output tokens
GPU cost
Lilypad markup
$0.00

Inference at cost

Use your trained model with zero Lilypad markup on inference.

Plans and training are billed separately. Per-token rates are published for each base model.

Trained on Lilypad
Your infrastructureadapter.safetensors

Own your weights

Export your LoRA adapter and use it with compatible infrastructure. No mandatory lock-in.

An adapter runs on top of its base model, under that model's license.

Your next model starts with a conversation.

Tell Lilypad what you want your AI to learn.