I pointed the agent at one of the hardest roles in AI hiring: a Machine Learning Scientist in generative audio — a real open req at an AI-native company in Cambridge, MA, pulled from its public careers page. The people who can train diffusion and transformer models on raw audio number in the low hundreds globally. Most sit at a handful of direct competitors. Almost none are looking.
If the agent could produce a credible sourcing strategy for this role, it can handle anything a normal hiring plan throws at it.
"This is one of the hardest roles to fill in all of AI hiring. Generative-audio ML scientists number in the low hundreds globally, cluster at a handful of direct competitors, and rarely move… You are not hiring from a market — you are recruiting ~150 named people. This is a named-account search, not a funnel. Posting and waiting will fail."
| Rank | Employer / lab | Fit | Sourcing note |
|---|---|---|---|
| 1 | iZotope / Native Instruments | Direct | Only local audio-ML research org actively hiring — people in motion |
| 2 | MERL (Mitsubishi Electric Research Labs) | Adjacent, elite | Small elite speech & audio group; pitch the jump to generative |
| 3 | Amazon / Alexa (Cambridge) | Adjacent | Biggest local speech bench; highest-volume crossover source |
| 4 | MIT (CSAIL, Media Lab) | Academic | Recruit via labs and papers, not job posts |
| Name | Profile (abbrev.) | Read |
|---|---|---|
| █████ ███████ | ML Scientist @ █████; PhD Audio & Music AI; diffusion/multimodal | ⭐ top fit |
| ██████ ██████████ | Audio + Music AI @ ██████; ex-███████ | strong |
| ██████ ███████ | Speech & Audio AI Scientist; MIT PhD | strong |
| ██████ ███████ | Speech & Audio Researcher @ ████ | strong |
12 more names in the full shortlist. The agent also flagged which candidates were already at the target company — and corrected its own first-pass undercount after discovering that narrow technical phrases ("music information retrieval," "neural audio") were the high-precision search keys.
It also returned comp benchmarks triangulated across the live posting, ZipRecruiter, Salary.com, and Levels.fyi; paste-ready Boolean strings in three tiers; "why now" outreach triggers (vesting cliffs, post-conference windows, competitor litigation stress); and the five actions the hiring manager should take that week.
Five chained Claude skills, each independently useful, sequenced by an orchestrator. Building blocks first, orchestration second.
- Talent Map
Market size, supply/demand, comp benchmarks, common backgrounds — sourced from live postings, Levels.fyi, funding data, and news.
- Target Company Builder
Ranked company list with funding, headcount trajectory, hiring signals, and why-they-fit rationale.
- Candidate Sourcing & Ranking
Ideal candidate profile + tiered Boolean strings; validated live against LinkedIn to converge on a named, deduped shortlist. Read the deep dive →
- Voice-Matched Outreach
First-touch and follow-up cadence written in my voice, trained on my own sent messages.
- Orchestrator
One command runs the chain end to end and returns the sourcing brief.
The agent flags estimates as estimates — pool sizes are triangulated and labeled as floors, with the precise count deferred to a LinkedIn Recruiter pull. Data gaps are flagged, not guessed.
Most sourcing starts with a job post and hope. This starts with the market: how many people can actually do the job, where they sit, what they cost, and what would make them move — before the first message goes out. It turns recruiting for a near-impossible role from a funnel problem into an account-based campaign with a named target list.
Want the full walkthrough? Email me — I'll run it live on your hardest role.