Dependency-free C++17 runtime and tiny learned sequence models (1-29 MiB) that run offline on ordinary CPUs, about a millisecond per small decision (0.8 to 1.5 ms on 8 threads of a desktop CPU), without GPUs.
[Raw Byte Stream] ──────── (Tokenizer-free / No Vocab Overhead)
│
▼
[Mamba-2 SSM Backbone] ──── (1.58-bit Ternary QAT / Linear O(N) Time)
│
▼
[Typed Decision Heads] ──── Output exact Byte Spans, Decisions & Scores (0 Hallucination)
▲
[C++17 SIMD Runtime] ──── AVX2 · ARM NEON · portable scalar (Zero External Dependencies)
Benchmark Results (from our 109-Page Technical Paper)
| Released model | Parameters | Weights file | One 64-byte decision (8 threads, desktop CPU) | Benchmark result | Baseline on the same benchmark |
|---|---|---|---|---|---|
| secrets-code | 1.9M | 0.98 MiB | 0.82 ms | F1 0.797 on CredData (164 repos): 3.5x the labeled credential lines found by gitleaks (3.1x on all 168) | gitleaks default rules: F1 0.337 (higher precision, 0.905 vs 0.879) |
| pii | 8.3M | 4.5 MiB | 0.89 ms | F2 0.769 on the PII Masking Benchmark | OpenAI Privacy Filter (1.4B parameters): F2 0.662 |
| secrets-bin | 54M | 28.8 MiB | 1.45 ms | 360 of 360 credentials on a synthetic probe cut from real binaries; 1 to 12 false alarms per set of clean binaries | strings + regex: recall 0.803 on the same probe |
- Standalone C++17 Engine: Zero external dependencies; SIMD kernels for AVX2 and ARM NEON plus a portable scalar kernel.
- Automated Verification: About 400 tests across the two repositories. On each platform, every kernel and thread count gives bit-identical results, and decisions are expected to agree across platforms.
- Honest limits: Confidence is not calibrated yet, and none of the three examples met every pre-registered criterion as first written. Both are documented in the paper.
- Code Repositories: purebyte-ai/purebyte (Runtime) · purebyte-ai/purebyte-train (Training Stack)
| AI Systems & Performance |
-00599C?style=flat-square&logo=c%2B%2B&logoColor=white){kind=icon}
|
| Distributed Backend & Cloud |
|
| Security & Systems |
|
- Indra Group (2023 - Present): Senior Software Engineer (Senior Specialist). Scalable backend microservices (Java 21, Kotlin, Quarkus, Kafka, Kubernetes) for high-criticality systems in defense and government programmes.
- Embention (2022 - 2023): Senior Software Engineer. Autopilot backend systems and real-time HIL/SIL flight simulation backends for Veronte autopilots, now used by Amazon Prime Air; customers in 70+ countries.
- Orizon (2021): Software Engineer. Performance optimisation of mainframe banking workloads for Spain's largest banks.
- Afterbanks Arcopay (2021): Software Engineer. PSD2 open banking aggregation platform (Java, Spring, MySQL).
- GESIO (2017 - 2021): Software Engineer. Led the 3-person eCommerce backend team (50+ customers) of an online ERP & POS SaaS with 3,000+ installations.
- BSc in Computer Engineering, Universitat Oberta de Catalunya, 2026 (GPA: 8.52 / 10).
- Final thesis: 9.9 / 10 with Honours (Matrícula de Honor).
- Top grades: Mathematical Analysis (10/10), Artificial Intelligence (9.7), Cryptography (9.7), Network Security (9.7), Component & Distributed Systems Engineering (9.4), Software Development Project (9.1).
- Red Hat: Cloud-native Microservices Development with Quarkus (DO378, Feb 2025) · Credly Badge.
- Winner Santander Explorer UA 2019 & Campus & Technology Award: Silicon Valley tech immersion trip (San Francisco, UC Berkeley, Stanford mentors) for Wazime (ultrasonic near-field data transfer).
© 2026 Pablo Sirvent · sirvent.ai