Independent researcher · Valencia, Spain
AI safety research engineer working on agent control, evaluation methodology and black-box optimisation — making the behaviour of long-horizon LLM agents auditable, testable and falsifiable.
I am an independent researcher and software engineer based in Valencia, Spain. My work evolved from search, game AI and combinatorial optimisation to efficient neural architectures, and now to the control and evaluation of LLM agents. Previous roles include AI programmer for PCFútbol and engineering work at technology startups.
Current focus: architectures and instrumentation that make agent behaviour observable at runtime — working memory, calibration and audit trails — plus optimisation methods for settings where gradients are unavailable. I work falsification-first: hypotheses state their refutation conditions in advance, results are multi-seed with standard errors, and negative results are published, not buried.
Sovereign Operating Memory Architecture — a four-layer cognitive architecture for long-horizon agents: bounded working attention (L1), immutable episodic ledger (L2), sovereign persistent knowledge (L3) and environment grounding (L4), regulated by an explicit context-pressure signal (somatic pressure, Pm). Zero-shot governance: no fine-tuning, weights untouched.
Zeroth-order optimisation of deep networks without backpropagation, replacing isotropic gradient estimation with structured low-rank subspaces. In a 15,000-forward-pass MNIST benchmark (109K parameters) against tuned Strong MeZO and SPSA baselines: 65.90% vs 10.60% final accuracy, at 3.13× lower wall-clock time (64s vs 200s).
Parameter-efficient fine-tuning combining weight permutation (TSP / bipartite) with spectral core adaptation (DCT, FWHT, wavelets). Documented across 12 benchmark experiments (v2–v12), including recovery of block-wise NF4-quantised GPT-2 and cross-layer shared adapters as small as 224 parameters / 0.88 KB.
An O(N) architecture replacing quadratic softmax attention and dense FFNs with a chunkwise-parallel complex phase delta-rule memory and a learnable spectral substrate (FWHT + DCT + Haar wavelets), with an empirical scaling roadmap toward a 72M-parameter language baseline.
A study of 2D wavelet transforms for non-autoregressive language synthesis. After experiments up to 200M tokens and an independent joint audit, the one-shot hypothesis was withdrawn: a hard entropic floor (~337 PPL native) makes it non-competitive — a published negative result. The same codebase produced a positive finding: multi-token speculative decoding v2 with 2.23× fewer forward passes and a 1.17× net CPU speedup.
A framework testing the thesis that learning is attention over frozen substrates rather than weight sculpting: low-rank multiplicative gating, spectral bases (DCT/Walsh), complex phase mixers and conformal geometry — across 287+ documented prototypes. Highlights: complex delta-rule MQAR at 99.11% vs 51.69% for a real-valued DeltaNet under iso-float conditions; causal-phase nGPT language modelling at 5.35 PPL with 80% fewer parameters.
Chess endgame tablebase compression with geometric position encoding: 99.93% accuracy on 3-piece endgames at 79.7× compression versus the Syzygy format — a laboratory for representation claims against exact ground truth before generalising to language domains.
Also: pac-clasifier (purifying archetype supervised clustering), 36 public repositories in total — search, sorting, TSP and tooling.
Research practice built for self-deception resistance:
Invited seminar — Instituto VRAIN, Universitat Politècnica de València (September 2026): agent architectures, evaluation methodology and the causal chain of this research programme.