Open theblack box.Know your model.

Project Hermeneus
Introducing Project Hermeneus
Continue reading
§ 01The betDelphi Lab · the wager · MMXXVI
The pitch
Small Language Models.
Domain depth.
A fraction of the cost.
§ 02The gapA market without receipts · Q2 MMXXVI

Everyone is fine-tuning. No one can prove what changed. A market without receipts.

+0%

more spend per request since the Q2 tokenizer change — per-token price held flat, the bill did not.Anthropic changelog · 2026

$0/eng·mo

burned per engineer on coding agents. Uber emptied its 2026 AI budget in four months as adoption ran 32% → 84%.The Information · Uber CTO · Apr 2026

0.0%

annual growth as teams move to fine-tuned small models. The shift is irreversible — inference is 280× cheaper since 2022.Stanford AI Index · 2026

⟦  ⟧

ways to prove what a fine-tune changed. No mechanistic audit standard exists in production — the verification stack is empty.Industry survey · Q1 2026

§ 03OpenBuilding · in private

The wager

The instrument is built.The market is empty.

StageBuilding · in private
BasedNew York · NY
Open toPilots · hires · talks
Respond by48 hours
Fabián del Valle

Fabián del Valle

Founder · Interpretability
PreviouslySenior Associate AI/ML Engineer · a top financial institution
EducationM.S. · New York University
FocusMechanistic interpretability · production model audit
+

Founding research engineer

Hiring · NY hybrid

Ship the audit pipeline at production scale. Comfort with PyTorch internals, kernels, and writing for non-researchers. Equity heavy.

Apply

Every lab scales capability. We scale understanding.