Fernando Cuevas
I own the layer between raw data and the business.
I turn raw data into clean, modeled, tested datasets and metrics that analysts, BI tools, and stakeholders use directly. As sole owner of a daily 10M+ record platform for Fortune 500 brands, I built the fact/dimension models, the campaign-and-spend metrics layer, and the self-service wizard non-technical teams run themselves — with reconciliation and QA built into the workflow rather than bolted on. My edge is that I bring software discipline to analytics (version control, testing, modularity) and I ship production LLM enrichment when it earns its place. Five-plus years across Databricks, BigQuery, and AWS, and a background in applied mathematics, are where the instinct comes from — make it correct, make it reproducible, and never trust a black box silently.
Selected work
Most of the systems below were internal production platforms, so they can't be open-sourced — but the architecture and the decisions are what matter. Two projects are public and clickable, and Shelf Aware is written up in full with its architecture.
Daily orchestration platform
Sole owner, end to end, of the orchestration platform behind campaign, spend, and audience reporting. The core pipeline resolves a client and date range down through campaigns → creators → posts, merges the spend tables, and rolls up cost metrics — architected as one platform that grew to fold in audience analysis and benchmark (IMBv2) exports, with a self-service wizard so non-technical stakeholders can run it themselves. Runs daily on Databricks (distributed PySpark, Unity Catalog, Delta Lake) with rerunnable, failure-isolated stages.
LLM segmentation & sentiment
Built production LLM segmentation on Gemini via the Databricks Mosaic AI Foundation Model API — classifying ~65K posts a year across a two-year corpus by region, locations, competitor mentions, and sentiment. I set a logistic-regression baseline first, held to a 98% accuracy target, and replaced the manual segmentation process, with batch handling, rate-limit management, and validation between the six-month reruns. Built to generalize to any post or brand.
Classification: 12h → 2min
Migrated brand-and-product classification from a 12-hour manual process to a 2-minute distributed PySpark job — swapping a naive per-post keyword scan for a dictionary-based approach running NLP feature extraction across millions of influencer records per cycle, validated against the senior analyst's golden-set output with full sample comparison.
Audience & campaign analytics
Built the Databricks SQL analytics behind creator and campaign reporting: resolved alias influencer IDs, tiered audiences (VIP → MEGA → MACRO → MID → MICRO → NANO), and computed demographic distributions (gender, age, ethnicity, interests, geography) with Hamilton-style apportionment so rounded counts stay mathematically consistent with their totals. On the spend side, multi-currency normalization rolled up to cost KPIs (CPP, CPE, CPVV, CPSV, CPME).
Export orchestrator
Built a wizard-driven export orchestrator in Databricks that runs four export lanes — audience analytics, campaign & spend, and two benchmark exporters — from one control surface. Selections persist across reruns in a Unity Catalog state table, a sticky-widget pattern surfaces only the relevant step, and results ship straight to Google Sheets via service-account auth or as a self-contained ZIP payload — no engineer in the loop.
OCR Agent
A desktop agent that watches a folder for screenshots, runs Tesseract OCR, queries an LLM, and surfaces the response. A queue-based threading bridge keeps UI updates safe, with env-var configuration and typed error handling throughout.
LLM influencer segmenter
A two-stage pipeline that pairs embedding-based retrieval with an LLM verification pass to segment influencers, with Unity Catalog support and a real test suite behind it — built the way production data code should be, not as a notebook.
Shelf Aware
A pantry-and-recipe assistant built around one hard-won rule: never let an AI guess about something that can hurt someone. The interesting part isn't the model — it's the architecture that keeps the AI on a leash. Detection is barcode-first because a spike showed vision models can't reliably read packaging; every AI reading passes through a mandatory human confirmation gate before any downstream logic runs; and the safety-critical path (allergens) is deterministic while only preference scoring is probabilistic.
Barcode-first detection
A spike showed vision models misread packaging text often enough to be unsafe, so a barcode/UPC lookup is the primary path and the model is the fallback — not the reverse.
Mandatory confirmation gate
No AI reading flows downstream until a person confirms it. The system is built to never act on a guess silently — the gate is non-optional by design.
Deterministic where it counts
Allergen filtering is rule-based and deterministic because it's safety-critical; only the preference ranking — where being wrong is merely annoying — is allowed to be probabilistic.
OLTP / OLAP separation
The serving database stays lean; change-data-capture streams to a separate lakehouse so analytics load never competes with the app that users depend on.
Stack
The tools I've actually shipped with — analytics engineering first, then the platform and infra underneath it.
analytics engineering
languages
warehouses & platforms
data engineering
ai & ml
ops & tooling
bi & libraries
Experience
- Architected the end-to-end orchestration platform on Databricks — API extraction, Unity Catalog transformations, multi-stage enrichment, delivery — with a self-service wizard for non-technical stakeholders.
- Owned daily pipelines processing 10M+ records across tens of terabytes for dozens of Fortune 500 clients, supporting $5M+ in annual contract value.
- Migrated brand-and-product classification from 12h manual work to a 2-minute distributed PySpark job (millions of records/cycle), validated against the senior analyst's golden-set output.
- Built production LLM segmentation on Gemini (Databricks Mosaic AI) — ~65K posts/year over two years, against a logistic-regression baseline and a 98% accuracy target — replacing the manual process.
- Cut deliverable turnaround ~50% with rerunnable, failure-isolated pipelines, and built reconciliation checks that caught a production data gap a shipped deliverable had already run on.
- Technical escalation point for a 5-person analytics team on the hardest deliverables (incl. Samsung Unpacked launch tracking); mentored analysts through new patterns and runbooks.
- Powered "Personas," an LLM-and-embeddings system that generated client-specific demographic persona sets — ran the LLM execution and moved embeddings from Redshift to local compute for the final step.
- Cut BigQuery cost and runtime ~15–20% on 150TB+ recurring operations through clustering and partitioning strategies.
- Owned a monthly foot-traffic export for a national retail client — ranking locations by month-over-month traffic — automated with Alteryx, GCP, MySQL, and SQL Server.
- Built pipeline work on the team-owned, multi-source pipeline (Excel, Alteryx, Experian, BigQuery, GCS) behind a client-facing web app that drove $1M+ in ARR.
B.S. Applied Mathematics & Cryptography
John Jay College (CUNY) — completed in three years.
Alteryx Designer Core
Certified in Alteryx Designer for data prep and workflow automation.
Python
Certified in Python since 2021.
Let's talk.
Open to senior data engineering, applied AI, and MLOps roles in NYC — hybrid or remote. If you're building something that runs at scale and can't afford to be flaky, I'd like to hear about it.