Skip to content
Case 04 · 8 min read

Case 04 · Technical lead · Search · Hiring marketplace

AccountMakers — four years leading the team, and a twenty-second search moved into the index.

Role
Technical lead · about 12 engineers over four years
Stack
TypeScript · Node.js · Express · MongoDB · Elasticsearch · AWS Lambda · Next.js
Scale
455 API handlers · 36 Lambdas · ~66k candidate records
Result
< 2 s / search target, was ~20 s

A hiring marketplace for sales talent, where I led the engineering team for four years and rebuilt the search that its AI matcher now runs on.

~20 s to <2 s1
talent search
4552
API handlers
363
Lambda functions
~66k4
candidate records

01 — The problem

A marketplace with an ATS as its source of truth

AccountMakers matches employers with vetted sales and customer-success professionals: account executives, SDRs, customer-success managers and account managers. The recruiters behind it run everything through Bullhorn, their applicant tracking system, so Bullhorn is the system of record and the marketplace has to stay in sync with it.

Talent build a profile, verify their identity, get matched to jobs, interview and accept offers. Employers search for talent, save searches, get AI-ranked candidates for each job, message people in real time, schedule interviews, run background checks and pay placement fees. An admin API covers approvals, Bullhorn sync and matching runs.

02 — What I built

The foundations, then the team

I was technical lead from the start in 2022, for a team of about 12 engineers across four years. My own commit count is small, around 95 across the repositories, because most day-to-day feature work came from the engineers I directed. What I wrote myself tends to be scaffolding, infrastructure, integrations and one-off pieces the rest of the system depends on.

That includes the Bullhorn integration, the split into talent, employer and admin portals behind Cognito auth, and the first real-time messaging layer, which the team later grew into a full chat system. In 2026 I moved talent search from MongoDB to Elasticsearch.

The team built most of the product on top: the matching engine and its LLM re-ranker, embeddings, AI talent summaries, background and identity checks, invoicing, interviews, the employer-of-record integration for temporary assignments and timesheets, and most of the UI.

  1. Bullhorn OAuth service and the token-refresh Lambdas
  2. Talent signup with Bullhorn field mapping
  3. Invites, roles, teams and email templates
  4. Socket.IO messaging and the saved-search model job matching grew from
  5. Docker image, MongoDB backup Lambda, Stripe webhook
  6. Deploying the recruiter-reporting service with headless-Chrome PDFs

03 — The hard part

One refresh token, three environments

Bullhorn issues one rotating refresh token per API user. Every refresh invalidates the previous token. If development, staging and production each refreshed on their own, they would keep knocking each other offline.

So only one function refreshes. A production Lambda runs the full OAuth sequence — authorize, exchange the code for access and refresh tokens, then log in for a REST session token and URL — and stores the result. It then fans the fresh token out to the development and staging Lambdas, which only store it and never refresh. Each function has its own IAM role, and API middleware reads the stored token on every Bullhorn-backed request.

I prototyped the OAuth flow on its own in 2022, and the multi-environment version reached production in 2023. The refresher is triggered through EventBridge and SQS rather than by user traffic, so the shared session stays alive even when nobody is using the product.

05 — Built on top

The matcher my team built

That index became the retrieval layer for job matching. Two engineers on my team built the AI matcher on the search layer I designed, growing it from the saved-search model I had created earlier. I want to be exact about this: the matching code is theirs.

It runs as a funnel. Elasticsearch applies the hard filters. Rule-based scoring and OpenAI embeddings blend at 60/40 by default, shifting with how much signal each side has. An LLM then re-ranks only the top 100, in batches of 10. Candidates scoring under 50 or with thin profiles are skipped, and results are cached on hashes of the job and candidate data, so unchanged pairs are never paid for twice.

The prompt may use only the supplied data, excludes protected characteristics and must name at least one gap per candidate. Scores are capped at 84 when primary-skill coverage or title quality is weak. OpenAI is the default provider; Gemini is configured as a fallback, switched through configuration rather than automatically.

06 — What I'd change

Next

The search number on this page is a design estimate. I would put a latency benchmark on the live search, run before and after every change to the index, so the claim stays measured rather than estimated.

The provider fallback in the matcher is a configuration switch. I would make it automatic, with retries that move to the second provider when the first keeps failing, so a provider outage slows matching down instead of stopping it.

Notes

  1. ~20 s was the MongoDB aggregation; <2 s was the target. The Elasticsearch design is estimated at 400–500 ms — a design estimate, not a post-deploy benchmark.
  2. Across talent, employer and admin route modules.
  3. Schedules, integrations and auth triggers, mostly built by the team.
  4. About 12k approved, from the February 2026 search analysis.

Ask my portfolio