Independent AI researcher · Almeida Industries

I work on modular and self-expanding AI — growing capability by composing and routing over inductive biases rather than scaling one monolithic model — on a single consumer GPU, measured honestly.

The through-line across my work: generalization comes from the inductive bias, not from more compute over tokens. So the interesting question isn't "how big," it's which basis, composed how — and how you verify a claimed improvement without fooling yourself.

How I work: every experiment is pre-registered, run over multiple seeds, and reported with its real disposition — including the ones that didn't work. I publish the nulls as prominently as the wins, because a trustworthy map of what fits and what doesn't is the actual product.

• LLM-IMPROVER — can a harness improve any off-the-shelf LLM on a benchmark by verified means? A single-GPU study on ARC-AGI: benchmark as fitness function, a frozen verifier, held-out delta as the product. Finds that capacity and verified search are the real levers — and reports the four things that came back null.

• Grounding-Intervention — does causal understanding come from knowledge, or from acting? A 3B model's causal prior scores 0.86 from the variable names and collapses to 0.00 when you rename them; intervention recovers the true causes name-blind. Grounding comes from acting, not association.

• secure-pickle — a jar of pickles comes tamper-proof; so should your code. HMAC-verified deserialization that refuses to open a tampered pickle before loading it — closing the arbitrary-code-execution hole in pickle.load() (the one that ships malicious model weights). pip-installable, optional AES-256-GCM.

From-scratch MoE and transformer LLMs trained on a single 5090; a knowledge-graph tool over a 200k-document research corpus; work on compression, routing, and frozen-module composition. Write-ups landing here as they're packaged.

Small-scale, honestly-scoped studies on consumer hardware. The claims are real and bounded — the point isn't that they're large, it's that you can trust them.

_{The handle's a security joke that isn't one: pickle is remote-code-execution in a file, and safely loading untrusted model weights is a real ML supply-chain problem. I don't take claims — or model files — on trust.}