Traces are pre-computed by the shipped cascade. Nothing on this page calls a model.
This page only knows the five replies above. To classify your own replies, request access to the dashboard.
Undertow runs the post-reply sales lifecycle — intent, objection, delegation, scheduling, recovery — as a cost ladder. Classical rungs are the default path. The LLM is the escalation. The eval harness is the arbiter.
Why a ladder at all? Because the alternative is a furnace.
02 · the problem
The obvious design fires an LLM agent for every capability on every inbound reply — intent, tone, calendar, routing, objection, recovery. Six calls. It works in a demo. At volume it is a budget line nobody approves:
10,000 replies / day
× 6 agents
× ~2,000 tokens
─────────────────────
≈ 120,000,000 tokens / day
Tools in this space charge per seat or per send. The cascade means most replies resolve at the cheapest rung — your cost scales with complexity, not volume. Unit economics only close if the majority of traffic is resolved for nothing and the model budget is reserved for the residue that needs it. And every reply waits for the slowest agent to finish — latency is the twin of cost.
A cost ladder puts the majority in milliseconds at ~0 cost and keeps the LLM for the replies that are actually ambiguous. On the held-out set the residue that reaches the top rung is 33.3% — one reply in three, not six calls in every one. A rung-3 answer takes microseconds; the reply is answered before a human would have seen it.
Some intents must never depend on a model's mood. Unsubscribe, opt-out, bounce, auto-reply: rung 0, deterministic, compliance-safe by construction.
Here is that column with its rungs named.
03 · the cascade
Every rung answers with a label and a confidence. Below the gate, the reply climbs. At the gate, it stops — and the rungs above it are never woken. Cost and latency are consequences of that rule, not of a cache.
Traces are pre-computed by the shipped cascade. Nothing on this page calls a model.
This page only knows the five replies above. To classify your own replies, request access to the dashboard.
| rung | method | handles | cost | latency |
|---|---|---|---|---|
| [7] | Full reasoning — reads the whole thread | ambiguous, multi-intent, "not now vs never" | expensive | seconds |
| [6] | Meaning match — two models must agree | long-tail intents, no training data needed | moderate | ~21 ms |
| [5] | Trained classifier — 15 intents, agreement rule | subtle intent the linear model misses | GPU once per corpus | ms |
| [4] | Similar examples — all 5 nearest matches agree | paraphrase of things already seen | ~5 ms embed, CPU | ms |
| [3] | Pattern match — statistical classifier, gate 0.55 | the bulk of clear-intent replies | ~0 | µs |
| [2] | Keyword scan — exact phrases | "not interested", "remove me", "talk to X" | ~0 | µs |
| [0] | Header check — OOO headers, MIME, bounce codes | OOO, bounce, unsubscribe, calendar accept | ~0 | µs |
[7] full reasoning not reached · surface
[6] meaning match not reached
[5] trained classifier not reached
[4] similar examples not reached
[3] pattern match confirmation 0.72 · gate 0.55 ▲ resolved
[2] keyword scan no "not interested" · no "talk to" silent · pass
[0] header check no ooo header · no bounce silent · pass
action schedule · slot Mon 10:00 PT · phone captured cost ~0 llm calls 0Rung 3 emits a probability. 0.55 is the gate, and it was chosen on the cross-validated train residue — the 233 training items a model that had not seen them would have escalated — never on the test split. The test split is touched once, for the table below.
Rungs 5 and 6 do not get to overrule rung 3. They may only promote it: accept when their argmax equals the classifier's and their own score clears a floor. Raw entailment thresholds were measured first and refuted — 55% precision at the highest confidence. Agreement is what reached the bar.
Rung 5 is a classifier head fine-tuned on our corpus. It is the most expensive classical rung shipped, and it shipped because the measurement said so: the only lever after rung 4 that stayed at 100% precision while widening coverage. The ladder ranks options. The eval picks them. Expensive rungs are taken when measured to be needed — never reflexively avoided, never reflexively used.
Classical rungs are the default path. The LLM is the escalation. That sentence is the whole doctrine; everything else on this page is evidence for it.
That word — schedule — is not a classification. It's the next move in a pursuit that has an objective and does not stop until it's met.
04 · pursuit engine
A pursuit has one objective and does not stop until it is met — meeting_booked, referral_captured, objection_handled. Every reply enters the cascade, and the resulting action moves the pursuit. Silence moves it too. So does the clock.
Child pursuit inherits the parent's objective. When it closes its own, it closes the parent's.
meeting_booked closes on a confirmed slot. The invite is out; the confirmation is the last mile, and it is a human's "yes", not a model's.
A rule layer over the generic decision: is the objective still reachable with this contact? Alex's "not my call" fails it and the pursuit asks for a name instead of proposing a slot.
Availability window ∩ the prospect's phrasing, pure calendar math, .ics attached. Natural-language date parsing reads "Monday at 10am"; nothing reasons.
meeting_booked closes on a confirmed slot; referral_captured on an extracted contact. The child pursuit for Dana closes the parent's objective when it closes its own.
The tick loop runs on a clock. A pursuit in waiting past its window gets one recovery draft, then another, then a withdraw — state and clock, no model. A no-show is an operator signal that re-opens scheduling. Priscila Nunes at Tidewater went quiet for three days after "send me a case study"; her recovery is the one draft awaiting approval in the console below.
Autonomy is set per action, not per system. What goes out on its own and what waits for you live in the same view — "fully autonomous" is a slider you decide the position of. The system proposes; a human confirms where the stakes require it.
Day 1 was an objection. It got a playbook answer at rung 2, not an improvisation. Every specialist in the engine has a ladder of its own.
05 · specialist agents
Each agent is a specialist cascade with its own cost ladder. Most ship LLM-free — not as a constraint, but because the measurement closed the gap at a cheaper rung.
Alex's first reply — "Not interested. We're already on Salesforce and it's working fine for us." — hits the competitor keyword scan at rung 2 (already on, working fine). Type → angle is a direct mapping into the playbook; the angle rotates per thread so a prospect never reads the same reframe twice. Shipped: type at [2]–[3], angle at [1]. No LLM in the loop.
Natural-language date parsing + timezone from contact metadata + slot intersection against availability, .ics out. Rung 4 (LLM extraction) exists for "sometime after my trip" and has not been needed.
Regex for emails and phones, dependency patterns for "talk to X" / "she runs RevOps", a pronoun guard so "she" binds to Dana and not to the rep. Dana Ferreira was extracted from one sentence.
Sentence length, contraction rate, formality hits, greeting form. Pure statistics. Style mirroring is feature extraction, not reasoning — this ladder has an LLM rung on paper and it is marked likely never.
State + clock. Silence timers, no-show signals, attempt counts. Detection is temporal, not linguistic — no model is consulted to notice that someone went quiet.
The exception, and it is laddered too: templates first (confirmations, invites, acknowledgments), style-adapted templates second, and the LLM only for open replies — constrained by policy, style profile and playbook angle. The LLM is load-bearing here — and correct. Alex's whole thread shipped on rungs 1–2 of this ladder. The light in this column sits at the surface on purpose.
Every one of those rung numbers is a claim. Here is what backs them.
06 · the eval harness
Accuracy on classification is the wrong bar for a system that sends. The bar is acted precision: of the replies the cascade resolves on its own, how many got the right action. Coverage and escalation come second — they only count if precision holds.
| enabled through | acted precision | coverage | escalation to rung 7 |
|---|---|---|---|
| rung ≤ 3 | 25 / 25 | 33.3% | 66.7% |
| + rung 4 | 35 / 35 | 46.7% | 53.3% |
| + rung 5 | 47 / 47 | 62.7% | 37.3% |
| + rung 6 shipped | 50 / 50 by action49 / 50 by intent | 66.7% | 33.3% |
successive cascade stages, not a time series · lower is deeper · held-out n=75
The one miss by intent: "This really isn't something I'm involved in." Expected wrong_person; rung 3 and the meaning match both said not_interested at 0.93. Both intents map to withdraw, so the action was right and the email that went out would have been the same. Three other meaning-match models miss the same sentence under the same protocol — a ceiling of the method on this residue, not of the model.
| lever | measurement | verdict |
|---|---|---|
| word + char n-gram TF-IDF | raw 57/75 vs 54/75, but acted precision 91–94% at every gate; word-only 100% at 0.55. Char features flattened the probability distribution — better argmax, worse calibration | refuted · 2026-08-31 |
| per-intent gates, 5-seed repeated CV | 24/26 at +0 margin; never 100% at coverage ≥ the global gate's 25/25. ~20 out-of-fold samples per class underestimate the wrong tail | refuted · 2026-09-02code kept, OFF |
| rung 5 as a fine-tuned NLI head | best residue argmax of any judge (208/233) but agreement never above 97.4%; held-out 52/55 — three wrong accepts | refuted · 2026-09-03the 15-way head shipped instead (44/44) |
A rung is placed by a number, kept by a number, and removed by a number. The refuted rows are the same discipline pointed the other way.
Run the eval harness against any dataset. The protocol is the same one that produced these numbers. See every refuted experiment →
That table was measured on 303 training examples. It will be measured again on more — and the ladder gets cheaper each time, not more expensive.
07 · learning loop
When a draft goes out on assist, the operator's approve, edit or reject becomes a labeled example. The learning cycle appends it to the corpus, retrains the rung-3 model, and reports the before/after on the same harness that produced the table above. Approval rates per intent become autonomy hints: an action that has been approved unedited often enough is proposed for autopilot. The system becomes less dependent on the top rungs with use, because the rungs at the bottom are the ones that learn.
competitor_complement, autonomy assist.rung · angle · corrections.competitor → respond approval rate crosses the floor at n ≥ 5 → hint: eligible for autopilot.Labels from rungs 4, 5 and 6 never enter the corpus on their own. They are copies of exemplars or a judge's opinion, and training on them would teach the cheap rungs to agree with the expensive ones. They enter only with an explicit human correction. The expensive rungs are auditors, not teachers.
That's the whole system: a ladder, a state machine, a harness, a loop. Here is the screen it all lands on.
08 · the console
One screen. Every pursuit, the rung that resolved it, and the drafts waiting for a human. Nothing goes out that you cannot see here first.
Three planes underneath — knowledge, state, interface. Read the architecture →
09 · how you start
No install. No API keys to configure. Four steps, and the fourth one happens on its own.
Point Undertow at the mailbox your prospects reply to. Inbound replies start falling through the cascade immediately.
Start from the fifteen shipped objection types and thirty approved angles, or bring your own. Set availability for scheduling.
Every draft waits for your approve, edit or reject. Each verdict is a labeled example; the cheap rungs retrain on it.
When an action's approval rate clears the floor, the console proposes autopilot for it — per action, never per system. You decide.
10 · security
11 · pricing
Managed infrastructure. No API keys to juggle. The eval harness runs on every deploy. Start free, scale as replies grow.
For teams getting started with the cascade.
Most teams land here.
High-volume teams with full control.
On-demand for large-scale operations.
Cloud-hosted. Managed infrastructure. Every plan includes the eval harness, the learning loop, and the full agent suite. The cascade decides what runs where — you pay for volume, not for rungs. Overages are billed at the per-reply rate of your plan.