Sai Yashwanth

Builds AI agents — by his own account working on agents for due diligence since 2024, across two startups. Writes at https://medium.com/@theyashwanthsai and publishes code at https://github.com/theyashwanthsai.

---

Contribution to This Vault

• Jev - The Most Interesting Model Released This Year (2026-09-18) — a short opinion piece on Jev, arguing that a model which returns a distribution instead of prose preserves the uncertainty a generated answer discards.
• Jevals — an open-source evaluation framework released alongside it, using that model as the judge with confidence gating and escalation of ambiguous cases.

The piece carries no measurements; its interest is the framing and the tool. The proposal worth testing is that cheap scoring makes it practical to judge context — which memory to keep, which retrieved chunk to admit — rather than governing it with heuristics (Context Engineering, Clean Context).

Related Concepts

• System One Model · Calibrated Relevance Probability · LLM as Judge · Staged Judging · Context Engineering

Related Notes

• Jevals · Jev · TypeSafe
• Reception of Jev — where this piece sits in the wider reaction