Research Scientist resume example

• Owned a six-month research agenda on multi-step reasoning at OpenAI, choosing impactful research problems by comparing failure rates, available compute, and expected scientific value; pursued the long-running project autonomously through weekly experiment logs and stop/go reviews.
• Developed a machine learning technique for verifier-guided reasoning with a fixed token budget, comparing uniform sampling, uncertainty-based allocation, and a supervised baseline; contamination checks and component ablations isolated the allocation effect from the benefit of simply spending more compute.
• Built high-performance implementations of deep learning algorithms in PyTorch, profiling activation memory and throughput before changing batching and checkpointing; reduced a fixed evaluation run from 11.4 to 8.1 GPU-hours while preserving the same prompts, seeds, and quality checks.
• Collaborated with peers in evaluation and infrastructure to test generalizable ideas at large scale; ran the same baseline across three model sizes and rejected an apparent gain that disappeared on the held-out domain, preventing an unsupported full-scale run.
• Wrote the first-author project report for a verifier-budget study, testing new ideas for allocating extra reasoning steps to uncertain answers; ablations showed that selecting harder examples mattered more than simply generating longer answers, and independent reviewers reproduced the comparison from the run manifest.
• Assessed impacts of AI technology by testing confident errors and abstention in unfamiliar domains; reviewers approved a failure-case appendix and release limitation after benchmark accuracy failed to reduce unsupported answers reliably.
• Reproduced a baseline using pinned dependencies and held-out tasks, reported confidence intervals across seeds, and included failed ablations rather than only the strongest run.
• Owned a machine-learning ablation using fixed seeds and peer-reviewed assumptions, resolved a confounded baseline and completed a reproducible report of negative results.
• Built a high-performance implementation using profiled kernels, reduced evaluation runtime by 28% at equal accuracy and published limitations relevant to the research agenda.

• Reproduced an instruction-tuning study at Google Research using a versioned dataset, three random seeds, and a held-out domain split; found that one apparent gain came from duplicated examples and corrected the comparison before the team selected its next experiment.
• Built an instruction-tuning evaluation harness with dataset hashes, fixed decoding settings, and per-domain error labels; the research team reused the harness for the next study instead of rebuilding the comparison, and reviewers traced regressions to the changed training mixture.
• Investigated a discrepancy between a published metric and its source records, traced the transformation that changed the population, and corrected the calculation with a reproducible query.
• Compared the last successful data refresh with a failed run, separated missing source data from transformation errors, and reran only the affected interval. Kept the original query and corrected result together for review.

• Created a reproducible research project comparing reasoning accuracy and abstention on 1,200 synthetic multi-step tasks; published an internal project report with dataset-generation code, baselines, negative results, and a statement that synthetic-task gains do not establish real-world safety.
• Logged accuracy, generated tokens, and GPU time at three inference budgets on the same held-out tasks; extra compute helped harder reasoning tasks but did not fix confidently wrong answers, supporting a bounded adaptive budget with an explicit abstention rule.
• Generated synthetic source records with duplicates, late arrivals and corrected values; wrote assertions for row counts and key uniqueness, and recorded the expected effect of each case on the reported metric.
• Compared the analytical output with a manually calculated reference table, traced differences to a transformation step, and retained a data dictionary and rerun instructions alongside the corrected query.
• Owned the synthetic-data validation using a manually calculated reference, resolved duplicate-key inflation and completed a notebook that reproduces the corrected totals.

Relevant coursework: Machine learning, optimization, statistical inference, reproducible experiments; dissertation on model evaluation

• Published an independent methods note using reproducible queries, explaining the data grain, excluded records and sensitivity of the result to a changed denominator.