01 / 13

Delivery case · engineering talent search

A recruiter sent two boolean strings. We sent back two talent maps.

The ask arrived as raw search syntax and one line of intent: Bay Area, open-source contributors preferred. No volume, no seniority band, no definition of a good candidate. This is what we did with it — and what came back within a single working day.

21,713

candidates delivered

independent sourcing tracks

3,300

ranked for outreach

1 day

brief to handoff

02 / 13

The request, verbatim

What the client actually sent us

Not a spec — the two strings they had been pasting into search boxes, plus a hunch.

("Kafka" OR "streaming" OR "event-driven")
  AND ("distributed systems" OR "data platform")
  AND (Java OR Scala)

— for the Bay Area, San Jose
— someone who contributes to open source would be the ideal candidate

"Query processing" OR "query optimization" OR "optimizer" OR "compiler"
  OR "distributed database" OR "query engine" OR "OLAP" OR "OLTP"
  OR "transactional processing" OR "analytical processing"

— folks who contribute to Apache Flink, Trino, Spark, Hive are a great place to look

03 / 13

The problem

A boolean string tells you which words to match. It tells you nothing about who to hire.

Five questions the brief left open — each one changes the answer by thousands of people:

Open question	How we closed it
How wide is "Bay Area"?	21 normalized city patterns, not the literal string
Is open source required or preferred?	Hard gate on one track, scoring weight on the other — so we could measure the difference
Managers or individual contributors?	ICs ranked above managers; invertible with one filter
Should current employees appear?	Excluded before scoring; former employees kept and flagged
What if the right person uses none of those words?	The reason we ran a second, inverse track

04 / 13

Step one

Two strings describe two different engineers

So we tracked each candidate against both shapes independently. The client filters by shape, not by keyword luck.

Shape S — Streaming

Tools: Kafka, Flink, Pulsar, Kinesis, Beam, Spark Streaming, Samza
Concepts: distributed systems, real-time, CDC, exactly-once
Language: Java or Scala primary

Shape Q — Query engine

Tools: Trino, Presto, Hive, ClickHouse, DuckDB, Druid, Pinot, Calcite, Iceberg, Velox
Concepts: query optimization, execution engine, OLAP, OLTP, MPP, vectorization
Language: Java/Scala, C++ for engine internals

Applied to both

Geography: 21 city patterns
Exclusion: current employees of the hiring company
Hygiene: bots, org accounts, corrupt records removed

05 / 13

Step two

Judgement calls we made — and wrote down

Every one of these is reversible by the client. None of them is hidden inside a score.

Decision 01

Bay Area is 21 patterns, not one

Matching the literal phrase would have dropped roughly two-thirds of the relevant population. Geography is where most searches quietly lose their pool.

Decision 02

Open source is a weight, not a wall

The brief called contributors ideal, not mandatory. Apache contribution carries a +5 bonus and its own sheet. Gating both tracks on it would have hidden most of the market.

Decision 03

Managers score below ICs

The role targets people who commit code. Staff / Principal / Distinguished rank top; managers deliberately sit lower. One filter flips it if the client wants the leadership cut.

Decision 04

Boomerangs stay, flagged

Former employees keep a visible flag and a small penalty instead of being deleted. 176 of them in the larger track. Whether a boomerang is warm or closed is the client's call.

06 / 13

Track one · anchored on public code

Who is actually writing this code in the Bay Area

The hard direction: start from visible engineering output, then resolve who that person is professionally.

Bay Area developers scanned

145,624

—

Recent non-fork Java/Scala/C++ repo

2,092

1.4%

After excluding current employees

2,088

−4

Professional identity resolved

714

34%

Fully enriched and actionable

481

23%

Confirmed Apache contributors

golden

The 1.4% is the finding. Our own pre-run estimate was 20%. Owning a public repository and contributing to one are different things — a prolific Spark contributor may own nothing at all. We reported the gap rather than quietly re-cutting the funnel to look better.

07 / 13

The hardest part

A code handle is not a person you can email

Our first identity bridge resolved 273 of 2,088. We built two more independent channels and lifted it to 714 — a 3.3× gain on the same frozen pool, with the confidence of every match written onto the row.

Channel	Method	Confidence	Resolved
Canonical mapping	Direct code-host link on the professional record	high	270
Verified handle match	Same handle both sides, confirmed by name overlap + country	high, verified	407
Name + employer	First name + last name + current company cross-match	medium	37
Unresolved	Code identity only — reachable via repo, blog or handle	code-only	1,374

Confidence is never blended. A client who only trusts canonical matches filters to 270 and loses nothing. One who wants volume takes all 714 and knows exactly which 37 are the soft ones.

08 / 13

Track two · anchored on career history

Most data-platform engineers never ship public code

So we ran the inverse strategy in parallel: start from professional career history, overlay open-source signal where it exists.

19,625

candidates · 122 fields each · 77 distinct Bay Area employers

70.7% have an email

91.0% carry company funding context

26.1% full career history enriched

15.3% have a phone number

Company context on 91% of the pool. Funding total, last round, IPO status, category tags, momentum score. "Staff+ at a public competitor with vested equity" and "Staff+ at a high-momentum private company with unvested equity" are two different conversations — and two filters on the same sheet.

09 / 13

Prioritisation

Three ranked waves, not one undifferentiated pile

Tier tells you fit. Actionability tells you who to email on Monday morning — fit plus reachability plus enrichment depth plus employer context, in one rank.

Wave	Size	Composition	Use
Top A — most actionable	300	100% have email, 100% full career history, 99% company context, 78% currently at a named competitor, 21% at FAANG	Day-one outreach
Top B — strong wave	1,000	Next by actionability	Second wave
Top C — broader	2,000	Wider, still qualified	Volume campaigns
Reserve tier	3,611	Sampled and augmented, not yet enriched	Follow-up batches on demand

10 / 13

Why two tracks earned their cost

The two searches disagreed — and the disagreement is the product

Dimension	Code-anchored	Career-anchored
Pool delivered	2,088	19,625
Proof of hands-on skill	Direct — public commits	Inferred from title and history
Career depth	34% resolved	100% by construction
Email reach	27.1%	70.7%
Company funding context	—	91.0%
Found by both methods	26 candidates

Those 26 are the highest-conviction names in the entire delivery. They write public JVM code in the Bay Area and hold a matching senior data-platform role. No single-source search produces that set — it exists only because two independent searches agreed.

11 / 13

The handoff

What the client received

Not a CSV dump. A package that still makes sense six months later, opened by someone who was not in the kickoff call.

The data in three lossless formats. Multi-sheet workbook for people, flat CSV for the CRM, compressed line-delimited JSON for pipelines — the same full field set in all three, nothing trimmed between them.
Pre-cut segments. Bands and tiers, Shape S and Shape Q specialists, Apache contributors, has-email, has-phone, currently-at-competitor, Staff+, boomerangs, cross-track overlap — each a separate file.
A written delivery report. Selection criteria, the scoring model with point values, match-confidence breakdown, coverage statistics, and an honest caveats section.
The expansion map. Six named levers with the volume each adds, ranked by recall gained against precision lost.
The frozen source pool and the scripts. Every number reproducible from the artifacts shipped alongside it.

12 / 13

Going wider

"We need more" is a filter change, not a new project

Each lever below is a re-rank against the frozen pool. Priced in advance, at handoff.

Lever	Adds roughly	Cost to precision
Lower the score floor	Promotes reserve into actionable bands	none — re-rank only
Enrich the reserve tier	+3,600 with full career history	none
Widen recency 18 → 24/36 months	Engineers who changed languages since 2023	mild
Add Kotlin, Rust, Go	Client libraries, Rust query engines	mild
Scan 500–1,000 employers instead of 100	+5,000–10,000 matches	mild
Open geography to US tech hubs	Seattle, NYC, Boston, Austin, Denver	relocation required

13 / 13

What we say without being asked

The caveats page ships with every delivery

Quoted from the original report handed to the client:

The open-source funnel is narrower than it looks. 1.4% of Bay Area developers own a recent non-fork JVM repository — far below our own estimate. Contributing and owning are different things.

The lower bands are sampled, not enriched. They carry identity and signal, not full career history. A deliberate cost decision, stated on the row.

One contributor sweep timed out. On the career-anchored track the open-source flag comes from cross-track overlap only — 26 confirmed, not exhaustive. Marked deferred, not presented as complete.

Nineteen surnames are truncated at source. We kept them as they are. We do not invent characters, emails or phone numbers to make a coverage percentage look better.

Same brief, another role? Send the strings you already use, the geography, the companies to exclude, and one sentence on what "ideal" means. You get two independent candidate maps, ranked outreach waves, complete records, and the expansion levers priced up front.