Applied AI / ML Engineer

You will build the first version of hiring intelligence models, the system that turns recorded interview transcripts into evidence-backed candidate profiles and explainable, ranked shortlists that real recruiters act on. This is an applied ML role, not a research role. The center of gravity is LLM-based structured extraction done with engineering discipline, a transparent matching/scoring model, and the evaluation harness that proves (or disproves) that any of it works. You will own this pipeline end-to-end from raw transcript to number-in-database and ship it to production. We operate on one hard rule: LLMs extract evidence; they never make the final prediction. Every score the system produces must decompose to evidence a recruiter or auditor can inspect. If that constraint excites you rather than annoys you, you're our profile.

Responsibilities:

• Build the extraction pipeline: LLM-based structured extraction of competency evidence from interview transcripts, versioned prompts and rubrics, persisted outputs with provenance, evidence quotes, and measured accuracy against expert labels.
• Treated as an engineering system, not prompt tinkering.
• Build the v1 scoring/matching model: a transparent, requirement-weighted scoring model that ranks candidates against role requirements. Interpretability is a hard constraint. Lay the groundwork for learning-to-rank as recruiter-preference data accumulates.
• Build the evaluation harness: precision@K against recruiter picks, agreement with expert labels, calibration where used, and subgroup fairness checks on ranked outcomes. Every model version ships with an evaluation report.
• Run label operations: design and run the expert-labelling loop with our SMEs and the recruiter preference capture; own training-data snapshots and model versioning.
• Ship to production: Own the full lifecycle of prototyping, evaluation, deployment, and monitoring, working with Python, Postgres, S3 and standard job-queue/batch infrastructure.
• Work independently: take ambiguous problems, figure out the right approach, propose it in writing, and drive it to production with minimal hand-holding.

What Success Looks Like:

• 30 days: extraction pipeline producing versioned, evidence-linked outputs on real interview transcripts; evaluation scaffold running.
• 60 days: v1 requirement-weighted scoring model live behind a shadow run; expert-label loop in motion; first accuracy and fairness report.
• 90 days: a written, statistically honest read on model quality against recruiter picks, including what to keep, what to fix, and what to kill.

Requirements:

• 4-6 years of hands-on experience building and shipping ML/AI products to production and keeping them running. Not just research or notebooks.
• LLM-as-component discipline: you have built structured-extraction or LLM-pipeline systems with versioned prompts and measured accuracy; you treat LLM output as a data source to be evaluated, not an oracle.
• Classical ML fundamentals: linear/gradient-boosted models, regularization, honest train/test hygiene; you know what leaks and what overfits when you have only a few hundred labels.
• Evaluation literacy: you can design an experiment, argue precision@K vs AUC vs calibration, and you are willing to say "this doesn't work" when the numbers say so.
• Strong Python and the modern ML stack (PyTorch / scikit-learn / Hugging Face or equivalent), plus comfort with data plumbing (SQL, batch pipelines, Docker).
• A builder's mindset with proven independent ownership; you've shipped real products end to end.

Bonus Points:

• Learning-to-rank or recommender-system experience.
• Prior hiring-tech or assessment-tech exposure, or awareness of its regulatory landscape (NYC LL144 EU AI Act).
• Fairness/bias measurement experience on ranked or scored outcomes.
• Speech/ASR or multimodal (audio/video) feature-extraction experience: this is our roadmap, not the day-one job.
• Vector search / embedding-retrieval experience.
• Early-stage startup or founder-led environment experience.