I am an Applied AI Engineer at TeachScribe, completing an MSc in Artificial Intelligence Technology with Advanced Practice at Northumbria University London. With a BSc in Physics from the University of La Laguna, I approach AI through controlled experiments, measurement and explicit assumptions.
My work spans multimodal and on-device ML, efficient inference, ML evaluation and reliable AI systems. At TeachScribe, I designed, implemented, measured and delivered an Android image–text retrieval pipeline. Previously, I worked with Python and scientific communication at the Instituto de Astrofísica de Canarias (IAC).
|
|
|
Production-integrated applied ML · Android Benchmarked five vision-language models, selected TinyCLIP 39M/16 and deployed separate INT8 ONNX text/image encoders with native Kotlin preprocessing and RAM/SQLite embedding reuse. Retrieval stays on-device; teachers retain the final photo-selection decision. Measured on one Android tablet: 441.07 ms median / 459.28 ms P95 over 108 cached searches; 74.77% smaller text encoder after quantisation. Shipped within TeachScribe's 50+ UK-school product deployment (product reach, not measured feature adoption). |
ML notebook auditing · Deployed web application A Next.js/TypeScript tool that statically inspects Jupyter notebooks without executing their code. Deterministic checks flag potential data leakage, test-set reuse, threshold-tuning risks and reproducibility gaps. An optional AI examiner adds narrative guidance; PDF reports and production-container HTTP E2E tests support the workflow. Try the live demo → · Inspect code and tests → The free-tier demo may need to wake after inactivity. |
|
Multi-agent orchestration and reliability engineering A sanitised, interactive simulation of a multi-agent control plane: capability-aware routing, execution leases, heartbeat tracking, quota fencing, bounded recovery, independent verification and a human merge gate. The public demo explains architectural decisions; it is not live operational telemetry or a control interface. |
Hardware-aware speech inference A reproducible transcription workflow built on
|
- IAC Scientific Publication ↔ Press Note Matching — Reproducible English/Spanish scientific-text matching with RapidFuzz, evidence-based ranking and explicit human-review tiers; related to my IAC internship.
- Computational Physics — Numerical Methods — Scientific Python notebooks covering interpolation, numerical integration, ODE solvers and stochastic simulation.
- AIoT Smart Home — ESP32-S3 — Simulation-based embedded C++ controllers, deterministic safety logic, Wokwi demos and a bounded TinyML experiment.
- Vizro/Dash open-source investigation — Contributed reproduction and regression-isolation evidence for a tab-rendering issue; the maintainer's separate fix PR #1752 was subsequently merged.
- Scientific publications and research — IAC science-communication work and an English-language Physics thesis on ferroelectricity and ferroelasticity (9.4/10).
I design around explicit execution criteria, bounded failure recovery, inspectable evidence and clear human decision points. Reliability depends on how a system behaves when inputs, models or dependencies fall outside expected conditions.
My evaluation approach starts with a baseline and controlled comparisons. I select metrics that reflect the practical cost of failure and inspect individual errors alongside aggregate performance to understand where a system is reliable and where it needs improvement.
| Area | Tools |
|---|---|
| Machine learning and inference | Python · NumPy · SciPy · pandas · scikit-learn · ONNX Runtime · TinyCLIP |
| AI systems and application engineering | TypeScript · Next.js · React Native · Kotlin · SQLite · Docker · REST APIs · GitHub Actions |
| Speech inference and media | faster-whisper · CTranslate2 · FFmpeg · VAD · CPU / CUDA |
| Scientific computing and development | Jupyter · Matplotlib · SymPy · LaTeX · SQL · Git |
| Embedded systems and AIoT | C++ · ESP32-S3 · Wokwi · Blynk · Edge Impulse |
Research interests: Reliable and efficient multimodal systems · model compression and quantisation · on-device inference · retrieval reliability during model migration · evaluation methods that reveal failures hidden by aggregate metrics.
Alberto Reyes Castro
Applied AI Engineer · Physics Graduate · MSc AI Technology Candidate · London, UK
Email ·
Portfolio ·
LinkedIn ·
Kaggle