recent

• Auto-Apply

  Autonomous from the posting to the filled form. It never hits submit. You do.

  Cuts time per job application from 5 minutes to 45 seconds. I click a posting; everything after that runs on its own. It tailors the LaTeX resume to that description, renders it to one page, fills the form and the free-form questions, attaches a cover letter, and stops at the submit button. The resume opens in a diff view that highlights only what changed, and every field the agent fills gets stamped with the app's logo, so I check those and skip the rest. 300+ applications in two weeks.

  Python, FastAPI, LaTeX, browser-use, a Chrome extension, pytest

  GitHub

• Project Hydra

  Backend scale engineering where the numbers are measured, not guessed.

  Nothing is simulated. The harness brings a topology up as real Docker containers, drives it with k6, scrapes Prometheus, and returns throughput and p50/p95/p99 at each load level. A rigorous sweep runs 100 to 8,000 RPS and fits the Universal Scalability Law with confidence intervals. Every data point carries its provenance, so measured and projected never get confused for each other.

  Docker, k6, Prometheus, Universal Scalability Law, M/M/c queuing models

  GitHub

• Preference Falsification and LLM Sycophancy

  What LLM agents say in public versus what they say in private, measured directly.

  24 agents over 9 rounds and 7 conditions, each making a separate private and public call per round, so the gap is measured rather than inferred. Social pressure behaved as a threshold, not a gradient: falsification appeared only under concrete career threats, while no pressure and moderate pressure sat at the stochastic floor. A whistleblower planted at round five moved about 20% of agents to honest public expression. Written up as a preprint.

  Python, LangChain, Google Gemini, NetworkX, Pandas, Matplotlib, Flask

  Proposal · GitHub

For everything else, see my essays and projects.