Work
Production AI systems, a documented method, and the pipeline behind them.
Three client builds and the site you’re reading. Where a metric or screenshot is still pending, it says so — nothing here is invented.
RAG chatbot for a multi-tenant payments SaaS, built and validated, scheduled to deploy in a few months — hybrid retrieval, PII scrubbed before the model.
AI Engineer — sole AI engineer at IDEA; I designed, built and lead the production RAG platform powering two live assistants (HR and electrical-engineering-rules).
Production multi-agent system on Semantic Kernel — Text-to-SQL, document and SharePoint tools, permission-aware.
Personal portfolio built with Claude Design and Claude Code, on a free stack — exhibit A of the method.
Nothing shipped under that filter yet — the rest of the road is in Next up.
HYPOTHESIS: An app shaped around one learner’s real friction beats a generic app shaped around the average user’s assumed friction — built for how I actually study Vietnamese, integrating listening, reading, and active recall in a single flow.
EXISTS TODAY: Spec drafted — input modes, session structure, and progression model are written down; no code exists yet. Scheduled after LLMwiki vault reaches its first milestone.
FIRST MILESTONE: A working session loop — vocabulary card → listening clip → typed answer → spaced-repetition scheduling — covering [PLACEHOLDER — Guillermo to confirm: vocabulary scope]. Target: [PLACEHOLDER — Guillermo to confirm: target date].
PLANNED START: [PLACEHOLDER — Guillermo to confirm: planned start quarter]
HYPOTHESIS: A reader agent ingests YouTube videos and LinkedIn reposts into a structured wiki vault; a manager agent then updates, corrects, links, and prunes entries as the field moves. A maintained knowledge base beats an ever-growing bookmark pile.
EXISTS TODAY: Agent roles defined (reader, manager), tool list and vault schema drafted (claim, source, date, confidence, linked entries, superseded-by). No code written.
FIRST MILESTONE: Manager agent successfully correcting [PLACEHOLDER — Guillermo to confirm: number of seeded wiki entries] when fed a conflicting source — claim diff logged, old version retained, link added. Target: [PLACEHOLDER — Guillermo to confirm: target date].
PLANNED START: [PLACEHOLDER — Guillermo to confirm: planned start quarter]
HYPOTHESIS: A curated, version-controlled collection of the Claude skills, prompt harnesses, and evaluation patterns proven useful across real builds. Each entry is named, tested and reusable. Formalising what works once means every future project starts one level higher.
EXISTS TODAY: Concept and motivation are clear; nothing is written or collected yet. Scheduled to start once LLMwiki vault is past its first milestone, because LLMwiki will itself generate the first batch of harnesses worth collecting.
FIRST MILESTONE: [PLACEHOLDER — Guillermo to confirm: number of harnesses] documented, individually runnable harnesses extracted from the Gestamp, Zelebrix, and LLMwiki builds — each with a one-paragraph description, a worked example, and the failure it prevents. Target: [PLACEHOLDER — Guillermo to confirm: target date].
PLANNED START: [PLACEHOLDER — Guillermo to confirm: planned start quarter]