Sai Anirudh Siddi — Writing https://www.saianirudh.blog/blog Sai Anirudh Siddi — AI Engineer working on ML systems, LLMs, model evaluation and SaaS. Projects, writing, and experience. en-us Probing medical LLMs for overconfidence https://www.saianirudh.blog/blog/probing-medical-llms-for-overconfidence https://www.saianirudh.blog/blog/probing-medical-llms-for-overconfidence Training a persona-vector probe that reads overconfidence straight off a medical LLM's activations — and the two changes that moved AUROC from chance to 0.88. Thu, 28 May 2026 00:00:00 GMT From 0.50 to 0.88 AUROC with activation normalization https://www.saianirudh.blog/blog/auroc-from-half-to-088 https://www.saianirudh.blog/blog/auroc-from-half-to-088 A short note on why normalizing residual-stream activations before a linear probe is the single highest-leverage change for behavioral detection in LLMs. Mon, 04 May 2026 00:00:00 GMT Monitoring interpretability in production https://www.saianirudh.blog/blog/monitoring-interpretability-in-production https://www.saianirudh.blog/blog/monitoring-interpretability-in-production Interpretability isn't just an offline research tool — wiring feature activations into Arize Phoenix and Sentry turns it into a live safety signal. Sun, 12 Apr 2026 00:00:00 GMT What I learned building ExoSeeker https://www.saianirudh.blog/blog/what-i-learned-building-exoseeker https://www.saianirudh.blog/blog/what-i-learned-building-exoseeker Lessons from building an AI exoplanet-detection pipeline on NASA Kepler data that hit >90% accuracy and won Best Use of NASA Data. Wed, 15 Jan 2025 00:00:00 GMT