GTM engineering · Chicago · remote

GTM systems that compound.

We build the pipeline that replaces manual SDR and RevOps volume: enrichment, scoring, routing, outbound, signals and reporting, wired together and measurable. Every system below is a working demo, not a slide.

Book a systems teardown See the demos

SDR: sends the messages. Volume from people.
RevOps: keeps the CRM honest. Process and reporting.
GTM engineering: builds the machine that does both. Volume from systems.

One sentence: GTM engineering is the discipline of building the automated systems that find, qualify, contact and route buyers, so a small team runs at a large team's volume without the headcount.

What we build

Five systems, in this order.

The order is the point. Each system consumes what the one before it produces, so building reporting before enrichment measures garbage precisely. Every card links to a working demo.

1demo →
Enrichment pipeline
Domain in, qualified contactable record out. Qualify before you buy the email.
domains → scored records
Why first: Everything downstream scores and routes what enrichment produces. Bad input, bad machine.
2demo →
Scoring model
Weighted timing signals with decay. Fit says who could buy; timing says who is buying now.
records + signals → ranked list
Why #2: Decides who gets a message this week. Timing over fit.
3demo →
Reply routing
Classify every inbound reply, update the CRM, page the owner. Minutes, not mornings.
replies → CRM + Slack
Why #3: The first system that touches revenue. Minutes matter.
4demo →
Signal detection
Daily watch on job boards, funding, exec moves and ad libraries. Accounts fire; lists stop going stale.
feeds → fired accounts
Why #4: Feeds scoring with fresh timing instead of stale lists.
5demo →
Reporting layer
Reply, interested, meeting and pipeline by segment × persona. The aggregate hides the two cells that carry everything.
activity → segment × persona
Why #5: Tells you which segment × persona is actually working, so you can double down.

Interactive demos

The systems, running.

Everything below runs in your browser on labelled sample data. No API calls, no tracking. The logic is the same shape we ship into Clay, HubSpot, Salesforce and n8n; the data is invented so nothing here is a client claim.

01Demo data

Enrichment pipeline

Enter a domain. Watch it move from a broad source through a scrape, an evidence-only description and a scored ICP test, to the threshold. Email credits are spent only after it qualifies.

client-side · no calls made

Broad sourceFirmographic lookup
Waiting.
Site scrape4 pages, summarised
Waiting.
AI descriptionFrom evidence only
Waiting.
ICP scoreTests + disqualifiers
Waiting.
Threshold cutQualify at ≥ 70
Waiting.
Email waterfallOnly after qualify
Waiting.

Output record

No record yet. Enter a domain and run the pipeline.

Try a qualified one and a disqualified one to see the waterfall gate.

Credit burn

Same batch, two orders of operations. Email waterfalls cost roughly 2.3 credits a record; a scrape and a small-model classification cost about 0.08.

Domains in batch1,000

Qualify rate22%

Enrich, then filter2,380 cr

Qualify, then enrich586 cr

75% fewer credits, or 1,794 credits saved on this batch. The saving is the disqualify rate; it grows with a stricter ICP.

Illustrative credit prices. Swap for your provider's rate card.

02Demo data

Scoring model: timing, not just fit

Eight accounts, five signal types, explicit weights. Flip between ranking by fit and by timing, toggle decay on year-old signals, and see why three fresh signals beat a perfect firmographic match.

Signals older than 12 months decayWeight × 0.25

#AccountFitTimingTotal

Why this rank

Kestrel

kestrelhq.com

35% × fit 72 + 65% × timing 75 = 74

Funding round6mo ago
Series B, $34M
Fresh capital funds headcount and tooling. Budget exists for roughly two quarters.
+25
Hiring for buyer function18d ago
4 SDR + 2 AE roles posted
Open roles for the people who would use the product mean the function is growing right now.
+30
New executive1mo ago
New VP Sales (ex-Gong)
A new leader re-evaluates the stack in their first 90 days.
+20

The lesson

Harborstack is the best fit on paper (95/100) but has no fresh signal. Kestrel fits less well (72/100) but 3 things changed there recently, so it outranks on total score.

Fit tells you who could buy. Timing tells you who is buying this quarter. A rep working the fit-only list sends Harborstack a message nobody asked for. Decay keeps two-year-old funding rounds from pretending to be news.

03Demo data

Reply routing

Simulated replies are classified, written to the CRM and posted to the owner's channel as a webhook chain. Then drag the response-time slider to see what slow follow-up costs.

Inbound replies

Event log

0 events

No events yet.

Process a reply and the webhook chain appears here.

Nothing in #inbound-hot yet.

Time to first response

A positive reply is a hand raised. The routing above exists so a human answers in minutes, not the next morning. Drag to see what waiting costs.

Minutes until a human replies to a positive30 min

Illustrative curve. Relative close probability vs. time to first human response on a positive reply.

Relative close probability

86%

of a 5-minute response

Close rate on positives

20.6%

base 24% at 5 min

Deals lost per month at 40 positive replies

1.3

Each one waits 30 min at this setting.

04Demo data

Signal detection

A daily watchboard over ATS job boards, funding, exec changes and ad libraries. Each rule has a content test and a time window; accounts that pass both fire into the scoring model.

runs daily06:00 America/Chicago · last run 2026-09-01 · 142s · 16 items checked

Rule: 2+ buyer-function roles opened in 30 days. Sources: Greenhouse-style, Lever-style, Ashby-style.

Harborstack2d agocontent ✓≤ 30d ✓fired
SDR (x3) · AE (x3) · Sales Enablement Manager
Greenhouse-style board
Ferro Labs5d agocontent ✓≤ 30d ✓fired
Head of Sales Development
Ashby-style board
Tessellate10d agocontent ✕≤ 30d ✓
Senior Backend Engineer (x3)
Greenhouse-style board
LumenOps12d agocontent ✓≤ 30d ✓fired
Account Executive (x2), first sales hires
Lever-style board
Kestrel18d agocontent ✓≤ 30d ✓fired
Sales Development Representative (x4) · Enterprise AE (x2)
Greenhouse-style board

Accounts that fired

6 accounts

Each fired account is pushed to the scoring model with the rule id and date, so the timing score above updates before anyone writes a message. Accounts that fire two rules in the same week go to a human first.

05Demo data

Reporting layer

The same 90 days of outbound, two ways. The aggregate looks fine. Split it by segment × persona and two cells are carrying sixteen.

Last 90 days · demo data

Reply rate, all segments and personas

3.9%

Sent: 15,260
Replies: 590
Meetings: 90
Pipeline: $2.84M

This is the number in most weekly reports. It looks like a healthy program. It is actually two segment × persona pairs carrying sixteen that are losing money, and the aggregate cannot show you which.

Prompt lab

How we write prompts that hold at volume.

A prompt that works on ten companies is a demo. A prompt that gives the same answer on ten thousand, cites its evidence and refuses to invent is a system. The difference is mostly discipline.

Same task: score a company against an ICP.

ICP scoring prompt

VagueLooks fine on 10 companies. Falls apart on 10,000.

You are a sales research assistant. Read the company information below and decide whether this company is a good fit for our product. Our ideal customers are growing B2B SaaS companies with a sales team. Give a score from 1 to 10 and explain your reasoning.

Company info:
{{company_info}}

'Good fit' and 'growing' are not testable. Two runs give two answers.
No disqualifiers, so an agency with a sales team scores 7.
Nothing forbids invention. Missing data gets filled with plausible guesses.
Reasoning is free-form, so you cannot audit or aggregate it.
A 1 to 10 scale with no anchors clusters at 6 to 8.

Holds at volumeSame output shape every time. Auditable. Cheap to re-run.

You score companies against an ICP. You never invent facts.

Evidence you may use: the SOURCE block only. If a test cannot be verified from SOURCE, mark it "unknown" and award 0 points.

TESTS (award full points or 0):
1. b2b_software (20): sells software to businesses. Fail for services, agencies, hardware, consumer.
2. stage (15): Series B–D, OR 80–600 employees.
3. sales_team (20): ≥3 open or filled SDR/AE roles.
4. buyer (15): pricing or customers imply ACV ≥ $15k (enterprise tier, 'talk to sales', enterprise logos).
5. crm (10): HubSpot or Salesforce detected.
6. geo (10): HQ in US, CA, UK or EU.
7. outbound_motion (10): sequencing tool detected OR SDR roles open.

DISQUALIFIERS (any present ⇒ score = min(score, 30)):
- consumer product · agency/consultancy/staffing · <40 employees · no sales function · regulated life sciences · government

OUTPUT strict JSON:
{
  "tests": [{"id": "...", "pass": true|false|"unknown", "evidence": "<quote or paraphrase from SOURCE, ≤ 25 words>"}],
  "disqualifiers": ["..."],
  "score": <0–100>,
  "reason": "<≤ 40 words, must cite at least two specific facts>"
}

SOURCE:
{{source}}

Each test is binary with a written pass condition.
Disqualifiers cap the score, so no amount of fit rescues a bad company.
'Unknown scores zero' removes invention and makes silence visible.
Evidence is quoted, so a human can spot-check 50 records in 10 minutes.
Weights are explicit; the score distribution is calculable before you run it.

Symptom → fix

What a scoring prompt does wrong in production, and the change that fixes it.

Keeps bad companies

CauseNo disqualifiers, or disqualifiers phrased as negative points instead of caps.

FixAdd explicit disqualifiers that cap the score. Test against 20 known-bad records before scaling.

Drops good companies

CauseTests require evidence the scraper rarely finds (exact employee count, funding stage).

FixAllow alternative evidence per test ('Series B–D OR 80–600 employees'). Log 'unknown' rates per test and fix the scraper, not the prompt.

Generic reasons

CauseReason field is unconstrained and the model pads it.

FixCap the reason at 40 words and require at least two cited facts. Reject outputs that fail the check.

Score clustering

CauseUnanchored scale. The model hedges toward the middle.

FixReplace the scale with weighted binary tests. The score becomes arithmetic, not opinion.

Different answer on re-run

CauseVague criteria and temperature above zero.

FixBinary tests, temperature 0, and a golden set of 100 records that must score within ±5 on every deploy.

Bill grows faster than the list

CauseFrontier model reads every record, including the 80% that a keyword rule could reject.

FixCheap-model-first routing: a small model applies disqualifiers, the frontier model scores survivors only.

Cheap-model-first routing

A small model culls, the frontier model scores survivors. Same output quality on the records that matter.

Raw domains100% of records
10,000 records
Rules62% of records
Regex + firmographic filters. Free.
Small model28% of records
Applies disqualifiers. ~$0.0003 / record.
Frontier model28% of records
Full ICP scoring. ~$0.012 / record.
Qualified19% of records
Enriched and routed.

Frontier on everything

$120

Routed · 69% less

$37

Per 10,000 records, illustrative list prices. The qualified set is identical because the small model only applies disqualifiers it can verify.

Outcome patterns

What these systems tend to produce.

Industry patterns · not client claims

Automate ~80% of SDR workflow

Enrichment, scoring, first-touch drafting and reply triage run as systems. Humans handle positive replies and calls.

80–100

meetings per rep per month, reported in public patterns

Qualify before enrich
Signal-timed sequencing
Reply routing under 5 minutes

Caveat: Depends on a real ICP and clean domains. Volume without deliverability is just spam.

Dynamic content from signals

First lines and offers generated from the signal that fired (role opened, round closed, exec joined), not from a template.

2–3×

reply-rate lift over static templates, in public patterns

Signal → hook mapping
Persona-specific proof
Forbid invention in generated copy

Caveat: Lift collapses when the signal is stale. Decay rules matter.

Call language → outbound loop

Phrases buyers actually use on recorded calls are mined weekly and fed back into sequences and ICP tests.

Weekly

feedback cycle from calls to copy, instead of quarterly rewrites

Transcript mining
Objection clustering
Copy tests by segment

Caveat: Requires consent-compliant recording and a reporting layer that splits by segment.

Patterns drawn from public GTM engineering practice. These are illustrative, not client claims. Real results depend on list quality, offer and deliverability.

Commercial foundation

The judgement under the systems.

Automation multiplies whatever it is pointed at. These three things decide whether it multiplies pipeline or spam: how deep the ICP goes, whether the funnel maths supports the plan, and whether the mail lands.

ICP depth

Most ICPs stop at layer one. Reply rates live in layers three and four.

1
Firmographic
Industry, size band, stage, geography
Table stakes. Every vendor filters on this. It removes nothing your competitors keep.
2
Technographic
CRM, sequencing tool, data warehouse, ATS
Says whether integration is easy and whether the buyer has already bought the category.
3
Situational
Hiring, funding, exec change, expansion, contract renewals
Timing. This is where reply rates come from.
4
Disqualifiers
Agencies, regulated verticals, sub-40 headcount, no sales team
The cheapest lever. A written disqualifier list removes 40 to 60 percent of most bought lists.

Deliverability checklist

0/8

Tick what you have. Anything marked critical that is missing means the rest of this page is not worth building yet.

SPF recordcriticalOne record per domain, under 10 DNS lookups, ends in -all.
DKIM signingcritical2048-bit keys, rotated, one selector per sending platform.
DMARC policycriticalStart at p=none with rua reports, move to p=quarantine once alignment is clean.
Secondary sending domainscriticalNever send cold from the primary domain. Use lookalikes that redirect to the main site.
Warm-upNew mailboxes ramp from 5 to 30 sends a day over 3 weeks before any campaign.
Volume caps30 to 50 cold sends per mailbox per day. Volume comes from mailbox count, not per-box volume.
Address validationBounce rate under 2 percent. Drop catch-all addresses unless verified.
MonitoringDaily inbox placement test and blacklist check. Pause a domain automatically on a failure.

4 critical items missing. Fix before any campaign runs.

Funnel maths

Before building anything, check that the plan closes. Move any number and the rest recompute. Defaults are mid-range for cold outbound to a well-defined ICP. Replace with your numbers.

Reps (or rep-equivalents of automation)2

Cold sends per rep per month1,500

Reply rate4.5%

Positive share of replies28%

Meeting from positive70%

Opportunity from meeting45%

Close from opportunity22%

ACV$38K

Sends3,000
Replies135
Positive37.8
Meetings26.5
Opps11.9
Won2.6

Bars are square-root scaled so the small stages stay visible. Values are per month.

Pipeline created / month

$452K

Closed-won / month

$100K

26 meetings a month at 3,000 sends. Each 1-point gain in reply rate is worth $101K of monthly pipeline at these ratios.

Engagement options

Three ways to work with us.

Start with the teardown if you are unsure. It is the fastest way to find out whether your problem is list quality, timing, response speed or reporting, and each of those has a different fix.