ALBERTO REYES CASTRO / LONDON, UK

Applied AI Engineer

Engineering reliable, efficient AI systems.

Multimodal ML • On-Device Inference • Agent Systems • Model Evaluation

Physics-trained engineer building end-to-end AI systems around measurable behaviour, deployment constraints, reproducible evidence and explicit human decision points.

View selected work LinkedIn ↗ Email

CURRENT ROLE Applied AI Engineer · TeachScribe

On-device multimodal retrieval, efficient mobile inference and measured engineering evidence.

TECHNICAL FOCUS

Multimodal & On-Device ML
Reliable AI & Agent Systems
Model Evaluation & Efficient Inference

EDUCATION

MSc AI Technology with Advanced Practice
Northumbria University London · 2026–2028

BSc Physics
University of La Laguna · 2025

LANGUAGES

Spanish · Native
English · C1 / IELTS 7.0/9.0
Japanese · Intermediate (around JLPT N3 level) · preparing for the N3 exam

01 / PROFILE

Engineering AI that can be inspected, measured and trusted.

I build AI systems from model selection through deployment and validation, with particular interest in multimodal systems, efficient inference, agent reliability and evaluation.

At TeachScribe, I designed, built, measured and delivered a fully on-device multimodal photo-retrieval pipeline for Android. My broader portfolio spans ML audit tooling, multi-agent reliability, speech inference and scientific computing.

I am interested in AI/ML roles in London where model behaviour matters in the context of the system around it: latency, memory, state, failure recovery, reproducibility and human oversight.

02 / SELECTED ENGINEERING

Three projects to start with.

If you have one minute, these are the strongest signals of how I work.

01 Applied AI · Production pipeline

TEACHSCRIBE

On-Device Multimodal Photo Retrieval

Designed, built and evaluated an Android image–text retrieval pipeline using TinyCLIP 39M/16, separate INT8 ONNX encoders, native Kotlin preprocessing and persistent RAM + SQLite embedding reuse.

Five pretrained vision-language models were compared before deployment. The final ranking path stays fully local and keeps the teacher as the final decision-maker. The pipeline has been integrated into TeachScribe and shipped within its deployment to 50+ UK schools; that figure is deployment reach, not measured adoption of this specific feature.

TinyCLIPONNX RuntimeINT8AndroidKotlinSQLite

441.07 msmedian cached search

459.28 msP95 · 108 searches

−74.77%text encoder bytes

76.3 MiBidle PSS · lazy image session

Public Case Study ↗ Watch Engineering Overview ↗ TeachScribe ↗

02 Reliable AI systems · Public simulation

AGENT CONTROL CENTER

Multi-Agent Orchestration & Reliability

Interactive explanation of a resilient multi-agent control plane with capability-aware routing, execution leases, heartbeat tracking, bounded recovery, separate verification and a human merge gate.

Six authored scenarios make failure behaviour inspectable: normal routing, reviewer outage, quota fencing, lost heartbeats, bounded recovery and independent verification. The public interface is deliberately sanitized and simulated; it exposes no live telemetry, provider calls, prompts or operational controls.

Agent OrchestrationLeasesHeartbeatsFail-ClosedHuman Approval

01Eligibility before routing

02One owner per attempt

03Bounded recovery

04Separate verification

05Human authority

Interactive Demo ↗ Watch Orchestration Walkthrough ↗ GitHub Repo ↗

03 ML evaluation tooling

MODELHAWK

Static Technical Auditing for ML Notebooks

Next.js/TypeScript tooling that parses Jupyter notebooks without executing them and surfaces deterministic evidence about risky ML workflow patterns.

Checks cover test-set reuse, threshold tuning against held-out labels, preprocessing before cross-validation, leakage/resampling patterns, weak metrics, reproducibility gaps and unsupported model-selection claims. The deterministic audit remains the source of evidence; an optional examiner adds narrative guidance rather than replacing it.

TypeScriptStatic AnalysisML EvaluationDockerCIPDF Reporting

01Parse, do not execute

02Deterministic evidence

03Regression-tested rules

04Optional AI narrative

05Production HTTP E2E

Current Sample Audit ↗ Live Demo ↗ GitHub Repo ↗

ADDITIONAL ENGINEERING

Broader technical range.

Useful depth, without competing with the three projects above.

SPEECH AI

Adaptive Whisper Transcription Pipeline

Reusable faster-whisper/CTranslate2 pipeline with hardware-aware CPU/CUDA execution, FFmpeg preprocessing, SHA-256 caching, atomic outputs and bounded decoding fallbacks.

Pythonfaster-whisperCTranslate2FFmpeg

GitHub Repo ↗ Kaggle Demo ↗

SCIENTIFIC NLP

IAC Publication ↔ Press Note Matching

Offline-first workflow using multilingual normalisation, RapidFuzz similarity, exact-title evidence, scientific-term overlap and explicit manual-review tiers.

PythonNLPRapidFuzzReproducibility

GitHub Repo ↗ Kaggle Demo ↗

EMBEDDED AI / AIoT

AIoT Smart Home · ESP32-S3

Simulation-based two-controller ESP32-S3 smart-home prototype with deterministic safety logic, optional Blynk monitoring and an explicitly separated TinyML experimentation path.

ESP32-S3C++WokwiEdge Impulse

GitHub Repo ↗ Simulation ↗

SCIENTIFIC COMPUTING

Computational Physics · Numerical Methods

Scientific Python notebooks covering interpolation, quadrature, ODE solvers and stochastic simulation with physical applications and reproducible notebooks.

NumPySciPyMatplotlibSymPy

GitHub Repo ↗ Kaggle Demo ↗

03 / EXPERIENCE

Applied engineering + scientific communication.

Jul 2026–Present

TEACHSCRIBE · LONDON / HYBRID

Applied AI Engineer

current

Designed, built, measured and delivered an on-device multimodal photo-retrieval pipeline for Android. Work includes vision-language model benchmarking, INT8 ONNX deployment, native Kotlin preprocessing, persistent embedding reuse, retrieval-quality analysis and physical-device latency, memory, thermal and energy evaluation.

Mar–May 2025

INSTITUTO DE ASTROFÍSICA DE CANARIAS · LA LAGUNA

Python & Scientific Communication Intern

250 h · 9.2/10

Developed Python workflows for multilingual scientific-interview transcription and segmented subtitling using the internship's original speech-processing stack, including SpeechRecognition, plus a semi-automated publication-matching workflow. Authored and adapted scientific communication in English and Spanish and produced technical documentation in LaTeX.

04 / EDUCATION

Physics foundation, current MSc in AI.

Current · Advanced Practice Jan 2026–Jan 2028

MSc Artificial Intelligence Technology with Advanced Practice

Northumbria University London · London, UK

Distinction-level current performance.
Selected first-semester marks: 85/100 · 75/100 · 70/100.
The 85/100 assessment was an end-to-end machine-learning project on credit-card fraud detection.
Relevant modules: Programming for AI · Machine Learning on Cloud · Artificial Intelligence for IoT · Information Governance and Cyber Security.
Advanced Practice pathway adds an applied industry/start-up component to the MSc.

Physics · Research Sep 2018–Jul 2025

BSc Physics

University of La Laguna · Tenerife, Spain

Final degree grade: 7.13/10.
Final Degree Project completed and defended in English: 9.4/10.
Selected marks: General Relativity 10/10 · Signals & Images 9.9/10 · Differential Calculus 9.5/10 · Electronic Physics 9.0/10 · Statistical Physics 8.7/10.

Read Thesis ↗ Defense Slides ↗

05 / TECHNICAL STACK

Capabilities backed by the work above.

ML & Evaluation

Python · PyTorch · Hugging Face · scikit-learn · XGBoost · pandas · model benchmarking · retrieval metrics · imbalanced learning · error analysis

Multimodal & Efficient Inference

Vision-language models · TinyCLIP · image–text retrieval · ONNX Runtime · INT8 quantisation · mobile inference · latency/memory profiling · Android

Reliable AI & Agent Systems

Capability-aware routing · stateful workflows · leases · heartbeats · bounded recovery · fail-closed state · separate verification · human approval

Application & ML Engineering

TypeScript · React · Next.js · React Native · Expo · Kotlin · SQLite · Node.js · REST APIs · Docker · GitHub Actions · SQL

Speech & Local AI

faster-whisper · CTranslate2 · FFmpeg · VAD · CPU/CUDA inference · INT8 / FP16 execution · local caching and output validation

Scientific & Embedded Computing

NumPy · SciPy · Matplotlib · SymPy · Jupyter · LaTeX · numerical methods · ESP32-S3 · Embedded C++ · Wokwi · Edge Impulse

06 / OPEN SOURCE

Regression analysis that helped a maintainer reach the fix.

MCKINSEY · VIZRO / DASH

Tab-rendering regression analysis

Reproduced a first-render regression in Vizro with Dash 4.2.0, compared dependency versions and progressively reduced the failing configuration across Graph, Container, Tabs, text content and pure-Dash controls.

The experiments showed the issue was broader than Plotly graph sizing and materially narrowed the failing path. A Vizro maintainer explicitly credited the analysis with helping narrow the bug and work out a fix. The maintainer's PR #1752 was subsequently merged.

TRACEABLE EVIDENCE

1Minimal reproduction

2Version comparison

3Component isolation

4Maintainer root-cause work

5Merged upstream fix

My Regression Analysis ↗ Maintainer Fix PR #1752 ↗

07 / TECHNICAL WRITING & RESEARCH

Long-form evidence, not just project cards.

Canonical reports only; project repos remain the source of truth for project-specific evidence.

ENGINEERING CASE STUDY

TeachScribe — On-Device Multimodal Retrieval

Architecture, measured retrieval and device-performance evidence, validation outcomes, limitations, privacy boundaries and deployment trade-offs.

Read Engineering Report ↗ Evidence Repository ↗

BSc PHYSICS THESIS · 9.4/10

Ferroelectricity & Ferroelasticity

English-language research thesis on structural mechanisms in β′-RE₂(MoO₄)₃ compounds, combining crystallographic analysis and long-form scientific reporting.

Read Thesis ↗ Defense Slides ↗

08 / SCIENTIFIC COMMUNICATION

Published science communication.

Co-author · Astronomía Magazine · Issue 311

The Partial Solar Eclipse That Reminded Canary Islanders of the Value of Looking Skyward

Co-authored with Ariadna Mesa and Hugo Fernández during my internship at the Instituto de Astrofísica de Canarias. Published in the May 2025 issue of Astronomía, Spain's first newsstand magazine dedicated entirely to astronomy, astrophysics and space-science outreach, in continuous publication since 1985.

The feature explains the scientific and public-engagement significance of the 29 March 2025 partial solar eclipse in the Canary Islands and situates the event within IAC outreach activity.

311May 2025 issue 3co-authors IACinternship publication

Read Published Feature ↗ Official Issue ↗

IAC · ORIGINAL ARTICLE

The Partial Solar Eclipse of March 2025

Original Spanish-language outreach article explaining the event, visibility, timing and safe solar observation for a general audience.

Read IAC Article ↗

IAC NEWS · ENGLISH VERSION

WEAVE observations of SN 2023ixf

Produced the English-language version of the IAC article, adapting specialised astrophysical findings for an international audience.

Read IAC Article ↗

09 / SELECTED CREDENTIALS & TRAINING

Training with technical depth and verifiable scope.

Issued by the Government of the Canary Islands · 325 hours

Artificial Intelligence & Big Data Applied to the Healthcare Sector

Course 24-38/006757 · IFCD104 · in person · Sep–Dec 2025

Government-backed technical programme combining the 300-hour IFCD104 specialty with 25 hours of complementary professional training. Core modules covered programming/data/AI models, machine learning systems, Big Data, probability & statistics with R, and healthcare information technologies.

110 h Programming, data & AI 40 h Machine learning systems 60 h Big Data systems 60 h Probability & statistics with R 30 h Healthcare technologies

Topics include Python, NumPy/pandas/scikit-learn, preprocessing, clustering, supervised ML, TensorFlow, computer vision/NLP, Docker/APIs, ML lifecycle, distributed systems, AWS/Azure/GCP, hypothesis testing, anonymisation and health-data interoperability standards including HL7 FHIR, SNOMED CT, LOINC and openEHR.

View Public Certificate ↗ English Syllabus ↗

AWS · OCT 2026

AWS Educate Machine Learning Foundations

Verified AWS Training and Certification badge covering core machine-learning concepts and applying the ML pipeline to business problems.

Verified Badge ↗

GOOGLE CLOUD · NOV 2025

Innovating with Google Cloud Artificial Intelligence

Google Skills badge focused on how organisations can use AI and machine learning to transform business processes.

Verified Badge ↗

NASA · APR 2025

NASA Open Science 101

Verified NASA Open Science badge covering open data, open code, open results, tools/resources and the ethos of open science.

Verified Badge ↗ NASA Course Details ↗

HARVARD UNIVERSITY · CS50 · IN PROGRESS

CS50's Introduction to Databases with SQL

Current OpenCourseWare study in relational modelling, querying, joins, schema design, views, indexing, optimisation and database scaling.

Official Course ↗

10 / LANGUAGE CREDENTIALS

Verified language proficiency.

BRITISH COUNCIL · JUL 2025

IELTS Academic UKVI — Overall Band 7.0/9.0

CEFR C1-level result for academic and UKVI purposes.

View Public Certificate ↗

UNIVERSITY OF CAMBRIDGE · SEP 2022

Cambridge C1 Advanced

CEFR C1 qualification covering reading, writing, listening, speaking and use of English.

View Public Certificate ↗

11 / RECOMMENDATIONS

Independent references.

UNIVERSITY OF LA LAGUNA

Academic Recommendation

Signed recommendation from a Full Professor of Applied Physics highlighting analytical ability, autonomous learning, communication and computational-physics / Python strengths.

Read Signed Letter ↗

INSTITUTO DE ASTROFÍSICA DE CANARIAS (IAC)

Professional Recommendation

Public recommendation from my internship supervisor covering Python automation, transcription/subtitling, publication matching, LaTeX documentation and scientific communication.

Read Public Recommendation ↗

CONTACT / LONDON

Interested in AI systems that have to work outside the notebook.

Open to AI/ML opportunities involving multimodal systems, efficient inference, agent reliability, model evaluation and developer-facing technical tooling.

Email me LinkedIn ↗ GitHub ↗