Hi, I’m Gabriel — an AI/ML engineer who designs and ships production generative-AI and machine-learning systems. My focus is AI workflows that eliminate the boring work, so people can spend their time on what actually matters: business decisions and getting things done.
What I care about most is rigor over buzzwords. I’d rather right-size a model — classical ML or an LLM — to the cost, latency, and reliability a problem actually needs, then prove it works with real evaluation: benchmarking against human-verified files, LLM-as-judge, and post-generation citation checks. Most of my work runs on AWS behind REST APIs and event-driven pipelines.
Where AI genuinely shines is turning large amounts of unstructured data into useful information that would take a human far longer to produce.
Talking with non-technical and AI-savvy people alike, the question I hear most is “could AI solve this for us?” More often than not, the honest answer is no — and it’s worth remembering what already works: well-structured data, knowing whether the data is even in your database, or just a bit of simple statistics. And when AI is the right tool, right-sizing it is what keeps you from overspending on tokens.
If you’re building something where the answer has to be right, not just plausible, I’d love to talk.
Merchants Bonding is a surety-bonds company, where high accuracy is essential — with financial data, there’s no room for mistakes. These are the projects I’ve been proud to be part of there.
3 production workflows
Architected the end-to-end pipeline — OCR → Bedrock/Claude via AgentCore → Strands structured output → human-review APIs — now powering three production workflows across the business.
85% less manual entry
Captures key line items at 98% field-level accuracy, tracked in production against underwriter corrections. Scaled throughput from ~700 to ~1,200 statements/month; per the company’s annual report, the platform drove an 85% reduction in manual data entry and 50% growth in instant endorsements.
~60 pages each
Automated summaries of 400+ sixty-page contracts a month with page-and-quote citations, grounded by post-generation string-match verification against the source PDF.
2,500+ files / mo routed
Built an 8-class classifier — TF-IDF + one-class SVM for OOD detection + logistic regression — routing 2,500+ files a month. Chose classical ML over an LLM for cost, latency, and interpretability.
solo-built & led
Built a system matching 3,200+ underwriter-opinion documents to database records, scoring section similarity and flagging discrepancies (red/yellow/green) to establish a single source of truth.
built with LLM-as-judge
Built a RAG + tool-calling assistant over dashboards and live database metrics, letting staff find dashboards and pull numbers in plain English.
VP-level analysis
Presented analysis to VPs showing no correlation between financial data and claims, driving the pivot from a predictive model to an LLM-with-rubric system applying underwriter guidelines.
≈3M → 900K
Mentored a developer redesigning a quarterly-analysis LLM workflow into parallel task-specific agents — cutting token cost ~70% and eliminating threshold hallucinations: the classic failure where a rule like “this metric must be above 90 to be good” wrongly flags a score of 91 as “bad.”
IOWA WHISKEY
Built by a bourbon hunter, for bourbon hunters.
Coming soon to the App Store
A SwiftUI iOS app — with a suite of on-device AI features — on an all-serverless AWS backend I architected and shipped solo. A nightly ETL builds a live catalog of every whiskey sold in Iowa from four public state sources; a Step Functions Distributed Map fans per-bottle inventory and delivery scraping out across 20 workers; and a rarity scorer ranks bottles by their availability across the state.
// architecture · AWS · data flows source → device
// see it in action
01 — Search
Search nearby stores to see what's actually on the shelf, so you know it's there before you show up.
02 — Hunt
Pick the bottles you're after and your time budget, and get a personalized route through the stores most likely to have them.
03 — Scan
Aim your camera at a bottle on the shelf to recognize it and pull up the details on the spot.
04 — Prices
Snap a receipt to log what you paid and help the community track the freshest prices seen across Iowa.
05 — Raffle
Turn a bottle into a raffle listing with guided, prefilled details — no spreadsheet required.
06 — Draw day
A step-by-step helper walks you through draw day and picking winners, start to finish.
