Systems engineer, researcher and professor. Mexico.
I work on complex and opaque systems: I reconstruct how they behave internally and turn that understanding into independent, portable platforms.
A self-hosting systems programming language with native performance, explicit memory control and scripting-language immediacy. Compiles through LLVM or runs source files directly via eskiuc run.
Two compilers implement the language through different lowering strategies. The C++ reference compiler lowers to a hand-written intermediate representation with constant folding, constant propagation and dead code elimination passes before emitting LLVM IR. The self-hosted compiler, whose lexer, preprocessor, parser, type checker and code generator are written in Eskiu, lowers directly to LLVM IR and reproduces its own output through a three-stage bootstrap fixpoint.
The language provides bounded generics monomorphised at compile time, structural interfaces, sum types with exhaustive matching, operator overloading resolved by operand type, closures, async/await over futures and typed channels, atomics, explicit allocation with defer and errdefer, checked nullable pointers and direct C interoperability, plus a standard library covering collections, networking, regular expressions, URLs, UUIDs and time.
Backends for AArch64, x86-64 and 32-bit ARM, with prebuilt toolchains for macOS, Linux and Windows. An Eskiu kernel boots in QEMU without libc, while the hosted runtime serves HTTP/2 over TLS.
In production: rendering modules in the ViroCore AR engine migrated from C++ to Eskiu use about 85% less memory across iOS, Android, Meta Quest and visionOS. Also used for the INE QR decoder below. Porting the engine to a Nintendo 3DS required a 32-bit ARM backend, a hard-float ABI and a static relocation model, all delivered back into the language. MIT licensed.
Infrastructure for visual and spatial computing. Maintains and extends the ViroReact ecosystem (over 130,000 annual package downloads) across native Android, iOS and React Native, alongside in-house computer vision, SLAM tracking and a cloud-anchor stack for persistent cross-device AR sessions that does not depend on Google Cloud Anchors.
Recent work: a WASM/WebGL2 build of the C++ engine bringing 3D and AR to the browser on an in-house visual-inertial tracker; Apple Vision Pro support through a maintained React Native fork; an authoritative replicated-state protocol for shared AR sessions with ownership by claim, optimistic concurrency and framerate-independent interpolation; and room-scale persistent anchors with cross-device relocalisation.
The work sits one layer below AR applications, in the perception, persistence and rendering systems they run on.
A document-intelligence and biometric runtime, reconstructed by reverse engineering commercial identity-analysis engines and rebuilt clean-room with explicit, inspectable behaviour.
Covers template classification, perspective rectification, visual OCR with locale routing, ICAO MRZ parsing, barcode extraction, front/back fusion, capture-risk signals, face detection and landmarks, head pose, passive liveness, portable embeddings and portrait-to-selfie verification. Ships with ROC/AUC/EER threshold tooling and REST, Python and CLI interfaces over ONNX Runtime. Structural consistency, capture risk, facial similarity and liveness are kept deliberately separate from legal document authenticity.
Passive face liveness detection. A maintained fork of Minivision's Silent-Face-Anti-Spoofing, abandoned upstream since 2020, modernised and measured rather than advertised.
Adds static-shape ONNX export validated against PyTorch, including a closed-form BatchNorm1d fusion that no off-the-shelf simplifier performs and that both CoreML MLProgram and TFLite require; an MIT-licensed face detector replacing the non-redistributable model upstream shipped; an ISO/IEC 30107-3 evaluation harness; and a frozen regression reference enforced in CI, with a geometry tier that re-derives every crop box from arithmetic alone and a numerics tier holding softmax drift under 1e-4.
Publishes three previously undocumented defects in the model's behaviour, along with two candidate geometric corrections that were fitted, measured and rejected as worse than no correction at all.
Clean-room reimplementation of the seven-layer cryptographic pipeline embedded in Mexican voter credential QR codes: AES-256-CBC, chained 8,192-bit RSA, NEON/SIMD key derivation and proprietary binary packing.
Reconstructed by loading and instrumenting the original closed ARM64 library under an emulator with a custom ELF loader and environment stubs, intercepting native cryptographic operations and capturing intermediate state. The result drops the native dependency entirely and is ported to C, Python, WebAssembly, Kotlin, Swift and Eskiu, every path producing bit-identical output, with mobile SDKs, browser execution and HTTP services validated against the captured reference. Published under Apache 2.0 with no vendor binaries, no decompiled source and synthetic fixtures only.
Security analysis of Mexico's national digital-signature infrastructure, used by roughly 27 million people. Covers deprecated primitives, short salts, weak password-based key derivation, private-key protection, GPU-assisted attack feasibility and reproducible recovery benchmarks.
Chapter in a volume published by the Instituto Politécnico Nacional, with F. Santander Baños (UAEH) and J. H. Sossa Azuela (IPN, Centro de Investigación en Computación). Documents a production remote identity-validation platform built in Mexico: two verification routes, document and facial, each pairing two computer-vision models through an adaptive decision strategy rather than a simple average, with a graded certainty output that routes ambiguous cases to human review instead of forcing a binary verdict.
University professor in artificial intelligence, data structures, algorithms and systems programming. My emphasis is on how systems behave internally rather than on consuming existing abstractions.
Also a personal trainer, currently studying strength training, weightlifting, biomechanics, sports nutrition and training volume and frequency management. The method matches the engineering one: observe, measure, model, test, iterate.
Compilers Reverse Engineering Native Systems Computer Vision Spatial Computing LLVM ARM64 WebAssembly Biometrics AR/VR Security Research
Premio a la Innovación Gubernamental, Gobierno del Estado de Hidalgo
