Hi, I'm
AI Engineer and data strategist. I build agentic pipelines, LLM systems and the data layer underneath them — every output behind a validation gate. And when the result needs to be seen, I produce the film, catalogue and campaign work that shows it.
Work delivered for
Master in Analytics, Innovation & Technology — INCAE Business School. Based in Costa Rica, delivering across LATAM, Europe and the Middle East.
I've spent six years on both sides of that gap. At national scale: the Hacienda Digital analytics platform for Costa Rica's tax and customs transformation programme — incremental processing on Delta Lake that cut reporting latency from four hours to near real time.
On the engineering side: Python and SQL workflows that replaced manual data preparation, competitor-pricing pipelines feeding weekly commercial strategy, and anomaly-detection bots that alert stakeholders before anyone opens a dashboard.
And mostly, now, on the AI side: LLM agents with tool access, RAG copilots with human approval gates, and multi-agent pipelines with hard validation rules — because an AI recommendation nobody can audit is not a recommendation, it's a liability.
Engagements usually start with one broken workflow and end with a system that runs without anyone remembering it exists.
ELT pipelines that replace manual preparation and nightly full reloads. Incremental merges, readiness audits and quality gates so a bad batch never reaches the table — or the report built on top of it.
Power BI and Tableau built backwards from the decision, not forwards from the data model. Fewer charts, clearer thresholds, and an owner for every number on the page.
LLM assistants with real tool access, RAG copilots grounded in your own catalogue, and multi-agent pipelines where every output passes a deterministic validation gate before anyone calls it approved.
Anomaly-detection bots that page you before the dashboard does. Demand models and price optimisers that make the margin-versus-volume trade-off explicit instead of intuitive.
Corporate and product video from shoot to final cut — script, edit, colour, subtitles and delivery. The kind of film a client can put on their homepage the same week, not the same quarter.
Catalogue photography, campaign creative and social assets produced through a scripted pipeline instead of by hand — thousands of consistent, on-brand images with a human reviewing fidelity before anything ships.
An industrial distributor with 29 manufacturer lines and almost no usable visual assets. I rebuilt the whole visual layer — film, catalogue, campaigns and the pipeline that produces them.
A machining and industrial-supply business that could describe its capability perfectly in person — and not at all online. No film, no consistent product imagery, no campaign system.
Same operating principle as the data work: generate at scale, review with judgement, never ship what hasn't passed a human.
Discuss a production projectAI systems first — agentic pipelines, LLM tool use and RAG, each public, runnable and documented as a case study with the numbers from real runs, including what failed.
Three single-purpose agents hand structured data to each other — scrape competitor listings, enrich against internal sales in Delta Lake, generate a grounded pricing recommendation — behind a hard validation gate nothing bypasses.
Ask a business question in plain English. Claude picks from a registry of typed Python functions, chains the calls it needs across three separate datasets, and answers with grounded numbers — no SQL, no dashboard.
Natural-language search over an industrial product catalogue, with local embeddings, cited sources, and an approve/reject gate that writes every answer to an append-only audit trail before anyone relies on it.
The engine behind the Elyon catalogue above: product-aware configuration, AI image editing, logo and watermark rules, multi-format export, resumable runs, and a visual QC gate that routes defects back to the right stage instead of the top.
An XGBoost demand model over 20 engineered features, wrapped in a scipy optimiser that finds the price maximising margin — or revenue — and makes the trade-off between them impossible to ignore.
A pre-build readiness audit that writes a pass/fail verdict per batch, a pandera quality gate that quarantines bad data with reasons, and an idempotent Delta Lake merge that retires the nightly full reload.
Rolling z-score anomaly detection across six business KPIs, with configurable hard floors and ceilings, firing branded email and Slack alerts on a schedule — before anyone has opened a dashboard.
Quantile-scored recency, frequency and monetary segmentation across 32,000+ transactions, mapped to eight segments — each with the retention action attached, not just a label.
Every repository ships with a case-study README: the problem, the architecture, the real run numbers — and an honest note on what's synthetic.
All repositories on GitHubI write about what actually happens when you put AI into a business — the mechanics, the gates, the parts that break.
New posts land on LinkedIn first — data, AI systems and the production work behind them.
Follow on LinkedInNot a logo wall — this is what shows up in the repositories and the client work above.
Tell me the decision you're trying to make faster. If I'm not the right fit I'll say so in the first reply — and usually point you at what is.
Hi, I'm Klarissa.AI engineer & data strategist
AI agents, data pipelines and automation, built around your business decisions.
Work delivered for
From the AI pipeline to the creative work that shows the result.
CLIENT WORK / ELYONWatch the corporate film · 0:58 An AI catalogue production pipeline, plus corporate film and campaign creative.
Three agents turn competitor pricing and internal sales into grounded recommendations.
Here's what my work puts in your team's hands.
What I deliver: an incremental data pipeline and operational dashboards that keep data fresh without a full nightly reload.
What I deliver: a copilot that searches your catalogue, cites its sources and records human approval before anyone uses the answer.
What I deliver: a repeatable product-image pipeline with brand rules, multi-format exports and human visual review.
Tell me the decision you're trying to make faster. Let's start there.
Email meOr say hello on WhatsApp