diff --git a/PRODUCT.md b/PRODUCT.md
new file mode 100644
index 000000000..1a3f0030e
--- /dev/null
+++ b/PRODUCT.md
@@ -0,0 +1,78 @@
+# Product description
+
+This document describes the value proposition of this product.
+
+## The problem
+
+Machine learning is a hard to take advantage of for most startups. They either don't have the time or the knowhow
+to deploy ML models into production. This problem exists for multi-billion dollar enteprises, it's 10x true
+for small startups.
+
+Python ecosystem is also hard to manage. Common problems are dependency hell and Python version conflicts.
+Most of the time, engineers just want to train and deploy an algorithm; everything else is distraction.
+
+Data is kept in databases that are hard for ML algorithms to access: MySQL, Postgres, Dynamo, etc.
+The typical ML workflow is:
+
+1. export data to a warehouse (e.g. Snowflake) or S3 (CSVs),
+2. run a Python script that will train the model (while fighting through dependency hell),
+3. pickle the model and upload it to object storage,
+4. download and unpickle the model in production, behind an HTTP API,
+5. serve predictions in a microservice.
+
+By the time this workflow completes, the data is obsolete, the algorithm is wrong and the ML engineer
+is polishing their CV or considering farming as an alternative career path.
+
+## The solution
+
+Colocate data and machine learning together in one system, train the models online, and run predictions
+from the same system with a simple command. That system in our case is Postgres, because that's where most
+startups keep their data. Postgres happens to be highly extendable as well, which makes our job easier.
+
+The new workflow is now:
+
+1. define the data with a SQL query (i.e. a view),
+2. train an algorithm with a single command,
+3. serve predictions with a SQL query.
+
+No Python, no code of any kind really, no dependencies, no exports, imports, transforms,
+S3 permission issues, deploys or JSON/GraphQL; from prototype to production in about 5 minutes.
+
+Here is an example:
+
+#### Define the data with a SQL query
+
+```sql
+CREATE VIEW my_data AS
+    SELECT NOW() - created_at AS user_tenure,
+           age,
+           location,
+           total_purchases,
+    FROM users
+    CROSS JOIN LATERAL (
+        SELECT SUM(purchase_price) AS total_purchases FROM orders
+        WHERE user_id = users.id
+    );
+```
+
+#### Train the model
+
+The function `pgml.train` accepts three arguments:
+
+- the model name
+- the `y` column for the algorithm,
+- the algorithm to use, defaults to Linear Regression.
+
+```sql
+SELECT pgml.train('my_data', 'total_purchases');
+```
+
+#### Serve the model
+
+The model is ready for serving! Let's serve this via SQL again:
+
+```sql
+SELECT pgml.score('my_model_1', '2 years'::interval) AS likely_purchase_amount_based_on_tenure;
+```
+
+You can call this directly from your app, no special infrastructure required.