Explore the Unknown

Research what's out there. Build what isn't.

Time

5 min

Plain English in. Working analysis out. No engineer required.

Inputs

Any data

Your CRM, public datasets, open APIs. The agent works on whatever fits the question.

Trust

Cited

Every answer comes with the data it learned from. No hallucinations.

Runtime

24/7

Runs on every new event. Schedule it, share it, keep it.

The gap

Between the dashboards
and the chatbots.

Dashboards count. Chatbots guess. Most companies have data and questions, with nothing in between actually doing the work.

QUESTIONS · this week SIGNAL · 47s 0 / 3 answered 3 / 3 answered
  • When will MRR drop?

    unanswered 23% risk · 30d
  • Which leads convert?

    unanswered 12 above 80% score
  • What predicts churn?

    unanswered 4 features · 91% AUC
Unanswered

Your data has answers nobody's asked for.

Dashboards count things. Spreadsheets summarize them. The questions worth asking — patterns, predictions, classifications — never actually get run on your data.

You

Will Sarah K. churn this month?

ChatGPT just now

I don't have access to your customer data — I can only suggest general churn-prevention strategies.

no answer · generic
Signal 18s · trained on your data

87% likely · within 14 days. Based on declining sessions, 3 unresolved tickets, and a missed renewal email.

predicted · cited
Wrong tool

Chatbots can't see your data.

Generic LLMs are trained on the public web. Your business runs on what's private — sales pipeline, support tickets, product events. They write paragraphs; they can't tell you which customer cancels next.

DATABASE 8,127,432 rows
support_ticket_4729 untrained classified
event_log_88121 untrained embedded
sales_call_2104 untrained scored
user_survey_077 untrained clustered
support_ticket_4730 untrained classified
0 models trained 4 models trained idle
Sitting still

Years of data. Doing nothing.

Support tickets, product events, sales calls, internal docs. The signal is in there. Without an agent that can actually work with it, it's just storage.

What ships

A model. An interface.
An agent. One Machine.

Every build comes out the same shape: a model trained on your data, the interface your team uses, and the agent that keeps them both running. Here's what that looks like, for one of the most common ones — catching churn before it cancels.

01 The model
0.91 AUC
↗ +6 pts vs week one 3,201 accounts

Trained on your last 18 months of churn. Returns a risk score per account, daily. Re-trains weekly as new outcomes land. Drift detected before it bites.

02 The interface
Churn watch sorted: risk ↓
  • Acme Corp 87%
  • Globex 81%
  • Initech 64%
  • Hooli 32%

A panel your CS team opens daily. Click a row, see the reasons. Shareable URL, no Signal account, no training needed.

03 The agent
06:00 UTC daily
  • 06:00 refresh sources · 3.2s
  • 06:00 re-score 3,201 accounts
  • 06:01 12 crossed >70% · pinged #cs-alerts
  • 06:01 re-train scheduled · Sun

Wakes up every morning, scores every account, pings Slack on threshold crossings, re-trains on the schedule. The agent stays on the question after you've stopped asking.

Compiled into "Acme Churn Watch" · one Machine · runs forever
How it works

Connect once.
Predict forever.

Plug in your tools. Signal trains a neural network on your data — with one head for classification and one for prediction — and ships live analytics that update with every new event.

Hover any source or output to trace its path.

EMBEDATTNMLPstep 1247 · loss 0.034 ↓ · acc 94.2% ↑PREDICTION4m agoSarah Kimenterprise · acme corp87%CHURN RISKLikely to churn within 14 daysCLASSIFICATION94% confFAILDamagedpackagingship_847.jpgflagged 2m agoRouted to QC review queuePREDICTIONnext 7d$48,200↗ +12%vs last week · ±4%ALERT03:47 UTCPayment volume spike+340%vs 7d baselinePinged on-call · ack expected in 5m
What early teams ask

Questions that don't fit anywhere else.

Map every YC company that pivoted in the last 3 years.Build a tracker for $500k houses 30 minutes from Boulder.Find every Reddit comment about hyperloop from the past month.Pull active recalls for car seats sold at Target since 2022.Build me an interface to query my Stripe data by cohort.Summarize patent filings citing 'mixture of experts' since 2024.Map every YC company that pivoted in the last 3 years.Build a tracker for $500k houses 30 minutes from Boulder.Find every Reddit comment about hyperloop from the past month.Pull active recalls for car seats sold at Target since 2022.Build me an interface to query my Stripe data by cohort.Summarize patent filings citing 'mixture of experts' since 2024.
Watch arxiv for new diffusion + RLHF papers and email me weekly.Make a dashboard for our Linear board grouped by team velocity.List every restaurant in Austin that opened in 2025 with a chef profile.Build a daily digest of every PR merged into our infra repo.Find lawsuits naming any Y Combinator company as defendant.Geocode our customer list and show retention by ZIP.Watch arxiv for new diffusion + RLHF papers and email me weekly.Make a dashboard for our Linear board grouped by team velocity.List every restaurant in Austin that opened in 2025 with a chef profile.Build a daily digest of every PR merged into our infra repo.Find lawsuits naming any Y Combinator company as defendant.Geocode our customer list and show retention by ZIP.
Manifesto

Most software stops at the search bar.
We think the interesting work starts there.

The questions worth asking don't have a Wikipedia page. They span domains. They require sifting open data, dirty data, and live data all at once. They end with something built — a spreadsheet, an interface, a quiet alert at 6 AM when the world has changed.

Signal exists for that work. It listens carefully, pulls from anywhere, cites everything, and assembles whatever instrument the question demands. Then it leaves the instrument behind, so you can use it again.

Notebooks show the work. AutoML ships a model. Signal ships the instrument.

Tune in.

Changelog

What we've shipped.

  • Apr 2026

    Machines, public

    Builds compile into reusable Machines you can run, share, or schedule.

  • Mar 2026

    Citation pulldowns

    Every research artifact ships with click-to-source citations and a confidence indicator.

  • Feb 2026

    Vision in research

    Charts, photos, screenshots, and PDFs are first-class inputs for research runs.

  • Jan 2026

    Drive integration

    Connect Google Drive — research and builds can read your documents directly.

Common questions

Things teams ask before they ship.

  • Doesn't this just call GPT?

    No. Signal trains specialized ML models on your actual data — classifiers, predictors, forecasters. GPT writes paragraphs. Signal builds prediction systems that run every day on your business.

  • What about data security?

    Your data stays in your tenant. Models train and run in isolated sandboxes you control. We don't train shared foundation models on your data, and you can wipe everything in one click.

  • Do I need to know SQL or Python?

    No. Describe what you want in plain English. Signal handles the schema, the joins, the cleaning, the training, and the deployment.

  • How long does it take to train a model?

    Most models train in under 5 minutes. Forecast and anomaly models can run on a schedule and stay continuously updated as new data arrives.

  • What can I connect to?

    Postgres, Snowflake, BigQuery, Stripe, HubSpot, Salesforce, Mixpanel, Notion, Slack, Zendesk, Shopify, and 30+ more. CSV and Excel uploads work too. Custom APIs in 10 minutes.

  • What if my data is messy?

    Cleaning is part of training. Signal handles missing values, schema drift, duplicate keys, and weird edge cases as it goes — and tells you what it cleaned so nothing is silent.

  • How much does it cost?

    Free during beta. No credit card. Production pricing rolls out at 1.0 — early-access pricing locks in for beta users.

Early alpha · Apply for access

Bring a question.
Take back a tool.

Apply for early access

Already on Signal? Sign in →