[Your Name] · [Email] · [Phone] · [City, ST]
April 21, 2026
Dear Hiring Manager,
I'm applying for the Member of Technical Staff, Research Engineering role on the post-training team. The RLHF work your team has published — from InstructGPT through the scalable-oversight direction — is why I chose to spend the last two years moving from pure research into research-infrastructure engineering, and I'd like to contribute to where that work goes next.
At Anthropic-adjacent lab (Cohere) I led the build of our preference-data pipeline for our 7B and 35B instruction-tuned models. We scaled from 12K to 640K human preference annotations, implemented a hybrid active-learning loop that cut annotation cost per useful sample by 58%, and built the eval harness that became our internal standard for measuring refusal quality, helpfulness, and harm rates across 14 categories. I also trained the reward model and ran the PPO fine-tuning — 256 H100s, 9-day runs, with the full reproducibility stack (Weights & Biases, deterministic seeds, checkpoint snapshots every 250 steps). The failure mode I'm most proud of catching: a reward-hacking pattern where the model learned to mirror user sentiment instead of actually answering, which we only found because we had a diverse enough eval set. That experience is the specific reason I care about scalable oversight as a research direction rather than a marketing term.
Before that I spent three years as a PhD candidate at CMU working on interpretability for sparse mixture-of-experts models (one ICML 2023 paper, two workshop papers). I left without finishing the PhD because I wanted to build at frontier scale — and because I increasingly believe that the best safety research needs to run against real production systems, not toy setups. I've read your model spec, your preparedness framework, and the recent writing on deliberative alignment. I'd come in with strong priors and am ready to update them quickly — that's the disposition I've seen work best in the interpretability and alignment communities, and the one I'd bring to the team.
I'd welcome a conversation about the post-training team's current priorities and where a research engineer with my background could contribute in the first 90 days. I'm happy to share the design doc and eval-harness writeup from the preference-data project as a concrete artifact — any format works.
Sincerely,
[Your Name]