Replay: illustrated answers

Every reply sorted and routed in 100 ms

Replies to an outbound campaign for a DORA compliance platform. Jev answers eleven fixed questions about each reply; rules in code turn the answers into a route. All names, companies and replies are invented.

ViaFromReplyRoute

Try your own reply

Select a reply.

Each reply carries a hand-written label: the category and the route a good SDR would give it. The numbers below compare the routing against those labels. In replay mode they only show how the rules behave; in live mode they are the measurement that decides the pilot.

Where the rules sent it versus the label

Category confusion (label down, Jev across)

Every mismatch, with the reason

Public benchmarks fill the model. Replace the assumptions with your own numbers; the pilot's first week is for exactly that.

Volume (benchmarks)

Cost of doing it by hand (assumptions)

Where the money is

The hours saved pay for the build, not much more. The case is speed and certainty: every positive reply reaches its owner within minutes, every opt-out is honoured the same second, and nothing waits for someone to open the inbox. HBR's audit of inbound leads put the average first response at 42 hours and found firms that responded within an hour were about seven times as likely to qualify the lead. That study is about inbound forms, not outbound replies, so treat it as direction, not as your number. The pilot measures your own.

At agency or platform scale the arithmetic changes: 20 clients on the same volumes mean about 6,600 replies a month, 330 hours of triage and still under $1 of Jev a month. That is the case for building it into a RevOps product rather than running it per team.

The pipeline

  1. Reply arrives from the sequencer's webhook. Its text goes into its own field, reply.text, apart from our campaign context.
  2. Hard rule in code. An explicit opt-out phrase suppresses the contact before any model runs. A compliance condition you can write exactly never depends on a model; Jev can only make it stricter.
  3. One request to Jev with the reply and eleven questions. They run in parallel inside the request, so the eleventh question costs almost nothing extra.
  4. Rules decide. Thresholds are set by what a wrong action costs: suppression needs 0.8, parking 0.7, a dated follow-up 0.6. Anything in between goes to a person.
  5. Action, with the raw answers and the rule that fired stored on the record.

The eleven questions, as sent

Thresholds (the part a human tunes)

What it does not do

  • It writes nothing. Replies to prospects stay with a person or a language model.
  • It does no arithmetic. Dates and deadlines are computed in code from what Jev extracts.
  • It does not know what is right for you. Thresholds are tuned on your own labelled replies.
  • Dutch works, slightly less accurately than English. Accuracy is tracked per language.
  • A probability is not proof. A 0.9 can be wrong, which is why the pilot measures against outcomes.

Run it live

Put jev_reply_server.py next to this file, set TYPESAFE_API_KEY as an environment variable (never in the page), run python jev_reply_server.py and open http://127.0.0.1:8788. The page detects the server and switches to live mode. The key stays on the server; the page only sends state and questions.