Autonomous growth agent·retail CRM·built in the open
Your CRM waits for instructions.
This one doesn’t.
GrowthOS reads your customer and order history, finds where revenue is leaking, writes the campaign to recover it, and tracks what actually converted. You approve the decisions — it does the work.
Upload your own CSVs, or start with 500 sample customers
Dormant VIPs
76 customers in your top spending decile have not ordered in 90 days. Their median basket is 2.4× the store average.
- Audience
- 76
- Recoverable
- ₹1.1L
- Confidence
- 0.82
A dashboard is a question. This is an answer.
Traditional CRMs hand you a query builder and assume you already know what to look for. Almost all of the work is in the knowing.
You guess which one matters this month
Surfaced by the agent, ranked by revenue at risk
You hand-assemble the filter
Already built, sized, and inspectable
You start from an empty box
Drafted per channel, refined in plain language
You open a report and interpret it
The funnel explains itself as events land
CSV in, revenue decisions out
Five stages, each one inspectable. Nothing here is a black box you have to take on faith.
- 01
Ingest
Customers and orders arrive as CSV. The importer validates, de-duplicates on a per-tenant key, and computes RFM scores and behavioural attributes for every customer.
csv → postgres · re-runnable - 02
Segment
A model reads the computed metrics and names the personas that actually exist in the data, rather than sorting customers into a fixed template of segments.
metrics → persona narrative - 03
Locate revenue
Deterministic rules surface seven opportunity types — dormant VIPs, churn risk, cross-sell, reactivation and others. Each one carries a sized audience and a revenue estimate.
7 rules · audience + estimate - 04
Draft the campaign
The agent writes channel-appropriate copy and three message variants for the chosen audience. You refine it in plain language until it sounds like you wrote it.
opportunity → copy + variants - 05
Send and measure
Approved campaigns leave through durable recipient jobs and a provider layer that reports back over signed webhooks. Duplicate and out-of-order events cannot move the funnel backward.
hmac webhooks → funnel
What the model is allowed to be wrong about
Money maths, eligibility and delivery state are arithmetic — they belong in code that can be tested. Language is the only thing handed to a model.
Decided by code
Written by the model
The interesting part isn’t the prompt
Anything can call a language model. The engineering is deciding what it is allowed to be wrong about, and containing it when it is.
- Three deployable services
- A Next.js frontend, an Express API, and a separate channel service that models a messaging provider over HTTP rather than an in-process function call — so retries, signatures and out-of-order callbacks are real problems the code has to solve.
- Typed boundaries around the model
- Every call runs prompt → JSON parse → Zod validate → database write, with a retry and a usable fallback. No free-form model output reaches the database or the UI.
- Tenancy enforced in the API
- Every database transaction carries the authenticated tenant into PostgreSQL. Row-level security and API ownership checks both reject cross-tenant access, including when a caller knows another tenant’s row id.
- Built for a free tier that sleeps
- The agent loop runs from an authenticated external scheduler, ingestion resumes after interruption, and durable recipient jobs retry provider startup and transient delivery failures.
Next.js 16 · React 19 · TypeScript · Tailwind 4 · Express 5 · PostgreSQL · Prisma 7 · Supabase Auth · BullMQ · Redis · Zod · OpenRouter · Vitest
See it find revenue in data
it has never seen
Create an account, then upload your own history or evaluate the pipeline on 500 sample customers and 3,000 orders — imported into your own workspace, never a shared sandbox.
Create an account→