kritwish.
  • Work
  • Systems
  • Playbook
  • Career

kritwish mondal, gtm engineer · diagrams sanitized: no customer data, no secrets

work / pipeline

Pillar 02 · Pipeline

Signals in, meetings out

Systems that notice who's ready to buy and get the right person in front of them, from a demo form or a cold list.

  • HubSpot
  • Salesforce
  • lemlist
  • Instantly
  • Slack
  • LangGraph
  • TypeScript

Why this had to be built

Only 1.7% of new contacts were ever marketing-qualified, inbound leads waited a median 44 hours for a first touch, and outbound had to reach thousands of local businesses without a list-building team.

Outbound program, best month
Reply rate
17% (adjusted)
Meetings booked
200+ in the month
New ARR from campaigns
$104K+ of $250K that month

Campaign-sourced deals, closed by AEs. Produced by the enrichment → campaign engine → lead scoring systems below.

Production

Inbound lead desk

  • HubSpot
  • Salesforce
  • Apex
  • Slack
  • lemlist
the deep dive →

What it is

Every inbound demo request is qualified, routed to the right rep by state, posted to Slack with the reason, and given a follow-up cadence in Salesforce.

What it did

  • Before: only 1.7% of 11,264 new contacts were ever marketing-qualified, so most never reached Salesforce.
  • Baseline: 222 inbound leads (Apr–Sep), 29.7% ever touched by a person, median first touch 44 hours.
  • Cadence tested at every layer: Apex 9/9, flows 93/93, end-to-end 42/42, real-data replay 101/101.
Before
29.7% touched · 44 h median first touch
Action
Qualify, route by state, Slack alert, Day 0/2/4 tasks
After
Being measured
Verified by
Real-data replay 101/101
Production

Signal tracking and hot accounts

  • Postgres
  • SQL
  • Firecrawl
  • SEC EDGAR
  • FullEnrich
  • Salesforce
  • lemlist
  • Slack
the deep dive →

What it is

One system that watches every target account for buying signals, keeps a research dossier on each, and turns the hottest ones into enriched, ready-to-work buyer lists every Monday.

What it did

  • Signals (demo forms, job posts, news, web visits, LinkedIn) feed a score that decays over time, with editable weights, a preview and a change log.
  • A weekly research dossier on hundreds of target accounts: homepage, filings, hiring, web visits, and news with an LLM judging which stories actually matter.
  • Buyers are enriched through a waterfall of 20+ data vendors, with a 30-day cache and a credit ledger so nothing is paid for twice.
  • First Monday run: 17 accounts, 136 senior buyers, and 61 new mobile numbers, each checked in Salesforce afterwards.
Production

Lead enrichment

  • TypeScript
  • DuckDB
  • LLMs
  • Search and scrape APIs
the deep dive →

What it is

A multi-agent pipeline that turns a list of local beauty and wellness businesses into qualified, contact-ready leads: owner name, staff count, booking platform, SEO gaps.

What it did

  • Benchmark run (n=50): owner first name for 43 of 50, at about $0.0117 per lead.
  • Rewrote the Python original in TypeScript: 73 s mean per lead vs 240–320 s.
  • 50 of 50 at concurrency 40 with 0 errors, where the old version deadlocked at 10.
Before
240–320 s per lead · deadlocks at concurrency 10
Action
TypeScript rewrite: typed single-shot extractors, caching
After
73 s mean · 50/50 at concurrency 40
Verified by
Benchmark run, n=50, 0 errors
Production

Campaign engine

  • TypeScript
  • LLMs
  • Instantly
  • HubSpot
  • DuckDB
the deep dive →

What it is

Turns each enriched lead into a ready-to-load cold sequence: score the pain, segment by rules, write 5–7 emails and an SMS, pass a deliverability gate, export for the sequencer.

What it did

  • Four duplicated runners replaced by one plugin runner.
  • Segmentation, scoring, subjects and links are code; the LLM only writes prose.
  • In three of the four types, a lead whose copy still trips the spam check after retries is dropped.
  • Unsendable leads are filtered out before any LLM spend.
Production

SDR lead scoring and routing

  • NestJS
  • Postgres
  • BullMQ
  • Redis
  • HubSpot
  • Slack
the deep dive →

What it is

An event pipeline that turns every open, click, reply and bounce into a decaying score, then routes warm leads to an SDR who is actually working that day.

What it did

  • Bot opens filtered out; replies classified by an LLM.
  • Scores decay with a ~9.6-day half-life into five temperatures with 1 h / 4 h / 24 h SLAs.
  • Attendance-aware round-robin; hot leads can never be dropped.
  • Each assignment creates a HubSpot task with an AI-written call brief.
Production

AI voicemail detection for cold calling

  • Twilio Programmable Voice
  • OpenAI Whisper
  • OpenAI
  • CallHippo
  • CRM
the deep dive →

What it is

A system that checks, just before an SDR dials, whether a real person answers a number. It places a short verification call, transcribes the first few seconds, and has a small LLM decide human or voicemail. Only humans reach the SDR. I built it end to end.

What it did

  • SDRs were making about 150 dials a day, and only 15–20% were picked up by a person.
  • Each number gets a short verification call first, recording only the first few seconds. No answer is rejected at once; an answered call is transcribed with OpenAI Whisper, and a small LLM decides whether it was a person or a voicemail greeting.
  • SDRs are only routed leads with a high chance of being picked up: numbers a person just answered, at most 8 at a time, refilled one by one as the SDR dials.
  • Pickup rate rose to 45–50%.
Before
~150 dials a day · 15–20% picked up
Action
Verification call → Whisper transcript → small LLM: human or voicemail; queue of 8
After
45–50% picked up
Verified by
Dialer pickup rate (self-reported)
Next pillar →Measured AI

Want systems like these at your org?

Message me on LinkedIn