Poindexter Labs jobs and careers

About Poindexter Labs

Olympiad-grade data creation * Author math/physics/compsci problems and multi-step reasoning tasks (proofs, multi-hop, ToT, chain-of-thought). * Produce fully-worked solutions, hints, rubrics, and counter-examples; deliver in LaTeX + JSON/Parquet. Stepwise reasoning & annotation * Decompose solutions into verifiable steps; tag dependencies, error types, and alternate paths. * Build evaluator prompts and scoring functions for exact, rubric-based, and partial-credit grades. Model training & evaluation * Curate instruction-tuning sets and reasoning curricula; run small/medium scale fine-tunes. * Design benchmarks and leaderboards with reproducible scoring, leakage tests, and stats reports. Cheat-resistant talent vetting (for existing contributor pools) * Create **secure, variant-rich assessments** to evaluate math/logic ability and proof writing. * Techniques we use: Parallel forms & dynamic variants (same construct, different instances; per-candidate seeding). * Hidden gold & cross-item invariants to detect copy/paste or tool-overreliance. * Time-to-solve and process evidence (scratch reasoning, step logs, LaTeX diffs) reviewed by experts. * Style & consistency checks (statistical signals + human review) to flag likely AI-generated text. * Closed reference sets and controlled tool policies; honor-tests that require original derivations. * Output: ranked shortlists, skills profiles, red-flags, and targeted upskilling plans. Red-teaming & safety * Adversarial prompts for math/logic hallucinations, spec violations, and content safety; failure taxonomies and fixes. Data governance & security * GDPR-first operations (ICO registered), least-privilege ac