AI & Tech Solutions Lead
Narrativewave · Remote · Guadalajara, Mexico
Sep 2025 - Present
Own AI strategy, architecture and delivery across three connected products, and lead a team of five AI engineers and data scientists.
WaveAI · agentic copilot for renewable-energy operations
- Architected a bounded investigation loop on LangGraph and MCP: resolve the outcome, assess evidence gaps, select a typed tool, validate and authorize, then reassess within explicit step and token budgets.
- Designed separation of authority: the application supplies authorized records, the agent coordinates, and the user confirms every state change through preview, confirmation, revalidation, execution and a durable receipt.
- Built typed tools over assets, telemetry, events, workflows, analysis history, pipelines, documents and charts, plus RAG with semantic chunking, hybrid retrieval and re-ranking: bounded evidence access, never open database control.
- Implemented evidence-guided completion states (complete, partial, clarification, blocked) so an exhausted budget is never reported as success.
- Built the evaluation and observability layer: golden datasets from real investigations, LLM-as-judge, grounding and hallucination checks, prompt regression tests on every change, prompt-injection defense, and full run tracing and replay with Langfuse and Arize.
Smart Summary & Priority Tracker · prioritization at fleet scale
- Classify the full open event backlog daily, track age and energy exposure, and rank sites for Operations, Performance Engineering, Asset Management and Executives across fleets of 100K+ assets per customer.
- Surface a cross-team Daily Focus with Act now, Investigate today and Watch labels, plus event playbooks.
- Made ranking criteria explicit, traceable and configurable per customer; Tell Me More launches WaveAI with the site, period and evidence already in context.
- Deployed across the customer base; triage dropped from about an hour to minutes.
MCP-first onboarding · DS delivery
- Built an MCP server and skill exposing typed operations over a canonical client specification: derive the package, check signal coverage, explain blockers, plan, simulate or apply, and verify, with runs persisted for auditability.
- Normalized heterogeneous client inputs (hierarchy exports, tag lists, historian metadata) into platform mappings while preserving source names; ambiguity surfaces as explicit review items, never silent assumptions.
- Took customer onboarding from about three weeks to 3–4 days and turned business-rule changes into config iterations instead of application releases.
Contributions & outcomes
- Lead and mentor a team of five AI engineers and data scientists; set the AI technical roadmap, architecture standards, and evaluation and reliability practices adopted across the product.
- Run weekly technical sessions with 3+ enterprise clients and executive reviews, translating operational needs into scoped, reviewable work and prioritizing what the team builds next.
Tools in practice
- LangGraph
- MCP
- Typed tools
- Hybrid RAG
- Evaluation harnesses
- Guardrails
- Langfuse
- Arize