Builds · Claude

Building with Claude

Where I use Claude: the engineering layer of a product — architecture, data model, backend, processing logic, testing, security and observability — plus reproducible research with Claude Science.

How I work with it

I use Claude Code for the part of a project that decides whether it holds up: defining the domain, designing the database, wiring backend and frontend, building the processing logic, and covering it with tests, authentication, security review and observability. For research work, I use Claude Science to build pipelines, run experiments, write tests and document failures as carefully as successes.

Projects

  • YAIS Lab Scholarships

    In production50% · Role of Claude

    Scholarship portal for the YAIS ecosystem: it centralizes calls, applications, confirmations, monthly milestones and applicant follow-up for each cohort, instead of a single static landing page per program.

    Role of Claude

    Version 1 was built with Claude Code: information architecture, data model, database schema, application flows and the first working frontend and backend. It set the foundation the platform still runs on.

    This project continued with Codex in a second phase.

    What it does

    • Cohort-based model: each call has its own dates, scope, requirements and entry path.
    • Application, confirmation and monthly progress tracking in one system.
    • Access rules and authentication reviewed and hardened in the second phase.
    • Query and performance optimization for the applicant-facing flows.
    ReactTypeScriptPostgreSQLAuthRLS
    Visit the platform

  • YAIS Video Dubbing

    In production60% · Role of Claude

    A dubbing service connected to a published Chrome extension that gives Spanish-speaking learners access to educational video content originally produced in English.

    Role of Claude

    Claude Code covered the engineering behind the product: the dubbing pipeline and its processing logic, authentication and access control, security review, the automated test suite, end-to-end web verification, and the logging and observability needed to run the service reliably.

    The remaining 40% — landing page, admin panel and management experience — was built with Lovable.

    What it does

    • Dubbing pipeline that processes English educational video into Spanish audio.
    • Authentication, access control and security review across the service.
    • Automated testing and end-to-end verification of the web experience.
    • Logging and observability to monitor runs and diagnose failures.
    ReactTypeScriptChrome ExtensionAuthObservability
    Visit the platform

  • YAIS Aymara Research

    Prototype80% · Role of Claude

    Open, reproducible research in natural language processing for Aymara, a low-resource language spoken by around two million people in Bolivia, Peru and Chile.

    Role of Claude

    Claude Science handled pipeline implementation, experiment tracking, test coverage and reproducible exports, while the research questions, benchmark design and interpretation stayed under human direction.

    The remaining 20% involved manual validation planning and community outreach for native speakers.

    What it does

    • Custom benchmark of 30 questions over Aymara Wikipedia with measured retrieval accuracy.
    • Documented failures: from 1/10 to 7/10 across six iterations, with each limitation published.
    • Reproducible pipeline with 63 automated tests and a public research site.
    • Open methodology: data sources, synthetic pairs and validation rules are disclosed.
    Pythonsentence-transformersLaBSENLLB-200FAISSBM25TanStack StartRecharts
    Visit the platform

  • In silico ASO design for ABCA4

    Prototype85% · Role of Claude

    A reproducible computational pipeline that designs candidate antisense oligonucleotides (PMO chemistry) to block the aberrant pseudoexon caused by ABCA4 c.161-395G>A, linked to Stargardt disease type 1.

    Role of Claude

    Claude Science built the scoring pipeline, wrote the test suite, drafted documentation and organized an adversarial review panel to surface critical issues before any claim was made.

    The 15% human input covered chemistry scope decisions, reviewer selection and interpretation of the clinical limitations.

    What it does

    • 224 passing tests and a fully documented, reproducible Python pipeline.
    • Adversarial review by six independent reviewers with public known-issues list.
    • Clear evidence-level labels on every claim; no synthesized or tested ASOs.
    • Explores PMO chemistry against the 2'-MOE/PS literature precedent.
    PythonBioinformaticsPMO chemistryASO designTestingReproducible research
    Visit the platform

Want to ship faster with AI?

I use agentic tooling daily on production systems. Happy to walk through what works, what doesn't and what it costs.

Ask me anythingSee the projects