You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Robin Langer — formally-verified mathematics, cryptography & trustworthy ML
Mathematician (PhD) and software engineer. I build systems whose claims a computer can
check: machine-verified proofs in Lean 4, cryptography from first principles,
interpretable and epistemically-honest ML, and autonomous research agents whose
reasoning is adversarially verified. The through-line is rigour you can reproduce — no
result below asks to be taken on trust.
Much of the mathematics here is not a scatter of separate projects but one connected
research program in algebraic combinatorics and representation theory, increasingly
machine-checked in Lean 4:
Symmetric graphs ⟷ Hecke algebras. My thesis studies vertex- and arc-transitive
graphs through the group acting on them. Every such graph is a coset graph
Sab(G, H, D) (Sabidussi), whose connection set D is a union of H–H double
cosets — precisely a basis of the Hecke algebra ℋ(G,H) = End_G(ℂ[G/H]), my
doctoral area. Lorimer's theorem (arc-transitive ⟺ D a single double coset HaH)
is the smallest non-trivial Hecke case.
→ lean-sabidussi (formalized, 0 sorry),
symmetric-graphs (per-graph Hecke
algebras + example catalogue), hecke-algebras.
q-series, symmetric functions & positivity. The same algebra one q-deformation
over: symmetric functions and Macdonald polynomials
(haskell, puzzles),
cylindric partitions and Warnaar's positivity program
(cylindric-partitions,
warnaar-glue), with the umbral /
finite-operator calculus as the Hopf-algebraic backbone
(lean-hopf-algebras,
umbral-calculus).
Formalization as the method. The graph theorems and the umbral/Hopf structures are
written in Lean 4 and certified by #print axioms — claims a computer verifies,
not claims you take on trust. This is the through-line to the cryptography and
trustworthy-ML work: rigour you can check.
A parallel strand applies category theory to the compositional modelling of real
systems — directed containers with a Lean-verified spine
(macbeth-seed) and applied instances. Same
habit of mind: find the structure, then make it formal.
Worth sixty seconds — the heroes
Ordered by what a high-assurance reader scans for first: verification, then
cryptography, then trustworthy ML, then autonomous research, then range.
My PhD theorem (Sabidussi representation + Lorimer, for symmetric graphs) machine-checked in Lean 4 — 0 sorry; the headline theorems reduce to Lean's foundational axioms, verified by #print axioms.
A transformer learns the inverse RSK correspondence to 100% exact match (n=10) and 99.99% at n=15 — 0.00004% of the space seen, so algorithmic generalisation, not memorisation (a baseline MLP gets 3.07%). Paired with a formally-verified per-agent trust algebra.
A loop of amnesiac Claude agents — context cleared each round, notes the only memory — attacking an open positivity conjecture (Corteel–Dousse–Uncu / Warnaar). Across 27 agent instances it produced genuine partial results (a proof of the Q₁ case, a determinant identity for the Corteel–Welsh system, the Adjugate Monomial Theorem); the full conjecture remains open, and the repo says so.
A published GECCO 2026 result (categorical composition of GAs); a math-retrieval embedder that beats OpenAI's text-embedding-3-small (MRR 0.816 vs 0.461, reproducible benchmark); a real-time GPU Navier–Stokes smoke solver.
Expository writing index → — the prose behind the proofs, grouped
by what each piece demonstrates for a security reader: cryptography and formal
methods first, mathematical depth as the foundation.
And two program narratives — how the pieces fit together, not just what they are:
A loop of amnesiac AI agents attacking an open positivity conjecture (Warnaar / Corteel–Dousse–Uncu); produced new partial results (Q₁ case, Adjugate Monomial Theorem). Full conjecture still open.
A CLI for mail, calendar, contacts and messaging — the pa/wa tools.
PhD in mathematics (symmetric graphs and Hecke algebras). Trained models are published
as HuggingFace model cards:
math-embed (math retrieval, MRR 0.816),
torah-embed (Talmud retrieval),
rsk-transformer (inverse RSK, 100%
exact). Every headline result is verifiable — the Lean index tells you
exactly what is machine-checked and what is merely stated, and the
expository index is the prose behind it.
About
Categorized directory of all my GitHub projects — pinned overview