kritwish.
  • Work
  • Systems
  • Playbook
  • Career

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

← Pipeline

Production

Signal tracking and hot accounts

  • Postgres
  • SQL
  • Firecrawl
  • SEC EDGAR
  • FullEnrich
  • Salesforce
  • lemlist
  • Slack

The problem

We track hundreds of target accounts, and the signs that one is ready to buy come from many places: website visits, LinkedIn ads, news, job posts, new leaders. Even after I built a score, it lived in an internal console and an alerts channel, so the BD team still had no list, no reason to call, and often no phone number.

The first dry run showed the gap. Title keywords alone found the right people at only 9 of the 16 hot accounts (77 people), and up to 66 of the people the system would pick had no phone in Salesforce.

Constraints

  • Signals age at different speeds. A demo request goes stale fast. A leadership change stays true for a while.
  • Don't send BD accounts that are already being worked. Customers, open deals and accounts touched in the last 14 days are suppressed. In one snapshot that held back 92 accounts, 32 of which would otherwise have been hot.
  • Titles are messy. "VP Operations" at a health plan might own credentialing, or it might run claims.
  • Data costs credits, and reps' edits win. The system only fills blank fields and never overwrites a rep's value.
  • Timing. Numbers must be in Salesforce before BD reads the 19:00 IST post.

How it works

  1. Research every account weekly (the context layer). Each Monday a run gathers, per account: the homepage, SEC filings, news, open jobs and web visits. A name search for news returns namesakes, so an LLM judge labels each headline as about this company, only mentioning it, or a different company. Most steps stay fresh for 6 days, so a retry in the same week costs nothing.
  2. Turn facts into signals. The signal types are demo or form submits, high-intent and engaged web visits, LinkedIn ad and organic engagement, major and minor news, leadership hires and hiring in our functions, and new leaders or team members at the account.
  3. Let each signal decay. A signal's strength is its worth × confidence × freshness, capped so no single signal maxes out the score. Freshness follows one of a few shapes:
    • Fades: web visits, forms and LinkedIn engagement start at full strength and drop in a straight line to zero over a window.
    • Holds, then fades: news, job posts and new leaders stay at full strength for a while, then fade.
    • Call-based and CRM signals that build toward a date, or only count next to another live signal, are specified but start switched off.
  4. Combine into heat. Each family of signals counts only once, using its strongest signal, so three different sources beat ten web visits. Each family then closes part of the remaining gap to 100. Accounts land in hot, warm or watch, and one strong signal raises an instant alert.
  5. Pick up to 10 senior buyers per hot account. Titles are matched against target and excluded phrases. Unclear senior titles go to an LLM judge, once per distinct title. Picks are ranked SVP, VP, Head, Senior Director, then Director, with no C-level and no managers.
  6. Find and fill through the enrichment waterfall. Empty slots are filled by people searches across every domain the account uses. Missing phones, emails and LinkedIn URLs then go through a vendor waterfall of 20+ data providers, tried in order until one answers. A 30-day cache means nobody is paid for twice in that window, and every row records the credits it spent.
  7. Deliver. Fill-blanks-only push to Salesforce, signal history on the Account, a weekly sequencer list with a six-step cadence that waits for the BDR on every email and LinkedIn step, and a Slack post that tags each account's owner and BDR.

Weekly research and raw signals feed a decaying heat score; hot accounts get up to ten senior buyers, who are found and enriched through a vendor waterfall and delivered to Salesforce, the sequencer and Slack.

From scattered signals to a Monday call list

Receipts

Before
Heat lived in a console; BD had no call list
Action
Monday job: score, pick up to 10 buyers each, enrich, push
After
17 accounts, 136 senior buyers, 61 new mobile numbers
Verified by
Each number checked in Salesforce after the push
  • First live run: 9 accounts got their full 10 people and 8 came up short. 98 Salesforce records carried the week's tag.
  • LLM title judge: 174 distinct titles judged in the first dry run for $0.0034.
  • Context layer sample: 20 health plans researched end to end in 2m 39s, raising 18 signals (10 hiring in target roles, 8 leadership hires).
  • Editable weights: signal worth and cutoffs live in the database. They are edited in the console with a preview and a change log, and every account is re-scored on the next run.

What broke, and what I changed

  • Wrong people. The first run picked 13 chief operating officers, a committee member and two managers, because a keyword was enough to get in. Now the level is checked on every pick, and searches ask for VP, Head and Director only.
  • Empty accounts. Three accounts got nobody. The search used one domain, which wasn't always the one the data vendor used, and one company wasn't in the vendor's data at all. It now tries every domain on file, then the company name with a strict employer match.
  • Bought phone numbers got lost. 10 phone numbers were bought and then skipped because the person had no email. Known records now get their phone by record ID.
  • Duplicates. The ID-first match lowercased Salesforce IDs, which never match, so some Contacts were duplicated as Leads. IDs now match exactly.
  • A count that trusted itself. The phone count came from the push's own report; it is now read back from Salesforce.
  • A silent "started" message. In the first context run, the Slack "started" message crashed between claiming the run and posting, so the retry stayed silent. Now everything is read before the claim.

Honest negative

The news judge agreed with a hand-labelled set on only 23 of 30 headlines. And the 30-day cache mostly saves rate-limit slots and time, not money, because the vendor already re-serves identical requests free for 3 months.

What I'd do next

  • Decide whether BD should see accounts reps touched recently. The 14-day suppression keeps 13 would-be-hot accounts off the list.
  • Add a no-retry window, so a person whose phone keeps missing isn't looked up again every week.
  • Switch on the call-insight, champion-move and positive-reply signals once their data is ready.

Want systems like these at your org?

Message me on LinkedIn