Case Study

How a Specialist Recruiting Firm Stopped Losing Deals to Slower Competitors

Published with Gigantes Group’s permission. Colossus is an internal tool built for their workflow, not a product CustomOps sells.

Company
Gigantes Group
Industry
Executive Search · Japan and Asia-Pacific
Focus
Recruiting intelligence platform
Engagement Type
Client engagement
Core Platforms
RecruitCRM, LinkedIn, Supabase, OpenAI

The Client

Gigantes Group is a specialist executive-search firm focused on engineering, operations and manufacturing leadership across Japan and Asia-Pacific. With a lean team of recruiters handling a high volume of mandates, efficiency and speed-to-shortlist directly determine revenue.

The Problem

Database blindness — paying twice for the same talent

Years of operation had accumulated thousands of candidate profiles in RecruitCRM. Yet when a new mandate arrived, recruiters defaulted to paid LinkedIn searches and enrichment tools — often re-discovering people the firm already owned.

RecruitCRM’s keyword search is exact-match. A candidate entered as “Head of Manufacturing Operations” never surfaced for a “VP of Plant Engineering” mandate, however good the fit. The firm was paying to re-buy knowledge it already had.

The first-mover gap — calling after the job was posted

Companies planning to hire a VP of Engineering do not begin on LinkedIn. The signal appears weeks or months earlier: a Series B, a new facility, a leadership departure. By the time a role is posted, five other firms have pitched.

There was no systematic way to monitor these signals at scale — just Google Alerts and manual news checks, which reliably produced late intelligence.

No AI-drafted outreach

RecruitCRM ships a sequencer, but it cannot generate personalised messages per candidate. Recruiters chose between generic templates, which perform poorly in a relationship-driven market, and writing every message by hand, which does not scale across active mandates.

The Solution: Colossus

Colossus is a recruiting intelligence platform integrated directly into the firm’s existing CRM and LinkedIn workflow. It has three modules that compose into a single motion.

Module 1 — The Candidate Ranker

A recruiter pastes the job description into the Ranker. Rather than extracting keywords, the system converts the whole brief into a 1,536-dimensional semantic vector and queries a unified candidate pool by cosine similarity, returning a leaderboard ordered by actual fit.

Two streams, one pool

A daily harvester searches LinkedIn for fresh talent using configurable industry watchwords, vectorising profiles automatically. In parallel, every existing RecruitCRM candidate is continuously indexed into the same vector store — work history, skills, current role and summary normalised into one text block.

Every result is source-tagged

[CRM] means the firm already owns the candidate — open them in RecruitCRM at no extra cost. [WEB] means newly discovered, and enrichment unlocks contact details before outreach.

What changed

The CRM stopped being a write-only archive.

  • Mandates that previously returned zero CRM matches now surface 10–15 ranked existing candidates
  • Time-to-shortlist fell from 2–3 days of manual searching to under 2 hours
  • Redundant spend on recruiter credits and enrichment for already-known candidates was eliminated

Module 2 — The Talent Scout

The Scout monitors the market continuously from two signal sources. It scans news for company and industry events — funding rounds, facility openings, leadership changes, M&A — and classifies each on a 1–10 urgency scale. Separately it watches live job listings, so a posting by a target company becomes a direct hiring signal.

Signal intelligence matrix

Signal
Score
Why it matters
New facility / plant opening
10
Immediate headcount need — call before the job is posted
Series B/C/D funding round
9
Capital injection typically precedes engineering scale-up by 60–90 days
Executive leadership hire
8
New leaders rebuild teams; congratulations sequences see 3× reply rates
M&A activity
7
Reorganisations create both BD opportunities and candidate availability
Mass layoffs at a competitor
5
Elite talent reaches the market before it appears on job boards

Signals scoring 9 or above trigger an immediate alert. Any signal can be turned into a scoped harvester profile in one click, seeding the Ranker before the company has opened a requisition.

What that changed

  • The team contacts target companies 4–8 weeks before a role is posted, against an industry norm of afterwards
  • High-urgency signals leave a time-stamped record of when intelligence arrived — useful evidence of market expertise
  • Congratulations follow-ups after a leadership hire became a process, not a matter of someone happening to see a post

Module 3 — The Custom Sequencer

Selected candidates are enrolled in a multi-step sequence spanning a configurable 14-day window: LinkedIn connection, email, LinkedIn message, soft email. The cadence is managed end to end, with AI-drafted messages per candidate rather than a shared template.

Sequence cadence

Day 1 — LinkedIn connection request
Day 3 — Email, routed through the recruiter’s own Gmail
Day 5 — LinkedIn message, if connected
Day 8 — Soft follow-up email

Reply detection runs continuously alongside every step, not at the end of the sequence.

Design decisions that mattered

The emergency brake

Webhooks watch the recruiter’s LinkedIn inbox in real time. The moment a candidate replies, all future steps are cancelled instantly. No recruiter has to remember to switch automation off — it stops itself.

Durable state, not cron

“Wait three days” is a first-class durable event rather than a scheduled job, so the system knows exactly where every candidate stands even across restarts and infrastructure events.

Human-likeness throttling

Sends are staggered within each batch at randomised intervals. Fifty connection requests do not go out at 09:00:00 — they go out across the morning, mirroring how a person actually works.

What Changed

Before
  • Exact-match keyword search that missed equivalent job titles
  • Paid re-discovery of candidates already in the CRM
  • Google Alerts and manual news checks
  • Outreach after the role was publicly posted
  • Generic templates, or messages written one at a time
  • Manual monitoring to avoid messaging someone who had replied
After
  • Semantic ranking across CRM and freshly harvested profiles
  • Owned candidates surfaced first, at no additional cost
  • Continuous signal monitoring, scored and alerted by urgency
  • Contact 4–8 weeks before the role is posted
  • AI-drafted messages per candidate, at sequence scale
  • Replies cancel the remaining sequence automatically

Technical Foundation

Layer
Technology
Application
SvelteKit 2 (Svelte 5 runes), deployed on AWS Lambda via SST v3
Database
Supabase (PostgreSQL + pgvector)
Semantic search
OpenAI text-embedding-3-small, cosine similarity via RPC
Workflow orchestration
Inngest (durable step functions, cron triggers)
LinkedIn
Unipile (messaging, connection requests, webhook reply detection)
Contact enrichment
ZoomInfo, behind a swappable enrichment interface
Market intelligence
Serper.dev (news and profile discovery)
CRM integration
RecruitCRM
Outreach email
Per-recruiter Gmail OAuth
Alerts
Telegram

Each external vendor sits behind a typed adapter, so swapping an enrichment provider or adding a signal source does not touch ranking, sequencing or cost tracking.

What this is not. Colossus is not a mass-market SaaS product, and not a blunt automation tool. It is built for one firm’s workflow, where relationship quality, timing and professional reputation are the competitive advantage. Every sequence has a brake, every enrichment call has a tracked cost, and every match shows its source — the system amplifies recruiter judgement rather than replacing it.

Sitting on a database you are paying to rediscover?

CustomOps builds internal tools that turn the data a team already owns into something it can actually act on — ranked, sourced and wired into the systems people work in every day.