Fabricio – Python, Ruby, Typescript

An automated content platform that turns thousands of scraped web sources into original published articles with zero human writing. An 8-stage LLM pipeline extracts claims, clusters stories, and writes information-gaining content — meaning articles that add something new instead of repeating what's already published. Each stage runs on the cheapest model that meets its quality bar, cutting cost per article to a fraction of an all-frontier-model setup.

• Built an 8-stage LLM pipeline that turns thousands of scraped web sources into original published articles: ingestion, claim extraction, semantic clustering, editorial synthesis, validation and post-production.
• Designed a hybrid model-routing architecture — self-hosted local models (Ollama on a dedicated GPU node) for extraction and classification, frontier APIs (Claude, DeepSeek) only where judgment and prose quality matter.
• Built a model evaluation harness with a golden-reference set and LLM-as-judge scoring; used it to detect a silently degrading pipeline stage and replace the model on evidence: chose DeepSeek as the cheapest option that matched the quality bar.
• Implemented semantic deduplication and clustering with sentence-transformer embeddings, pgvector and HDBSCAN.
• Made the whole pipeline idempotent and resumable: PostgreSQL as the state layer, unique constraints and upserts per stage, safe to re-run after any crash.
• Built the supporting infrastructure end to end: PostgreSQL/pgvector on Docker, automated daily telemetry via systemd on a Linux node, Telegram monitoring, and a Next.js front end on Vercel.

A white-label client-retention platform for mortgage lenders and loan officers. Borrowers receive monthly branded home value reports built from multiple valuation data providers, and the platform monitors a network of 2,000+ mortgage and real-estate sites to detect when a client starts shopping for a refinance or HELOC, alerting their loan officer in time to save the deal. Serving thousands of active users across multiple lenders. I owned the data layer the product depends on: multi-source ingestion, valuation data flows, and database performance.

• Optimized the PostgreSQL queries behind the value-report generation path--the platform's heaviest workload: 88% runtime and 85% memory reduction, through execution-plan analysis, indexing, and query restructuring.
• Rebuilt the lender data import system — the onboarding path for every new lender joining the white-label platform: turning a fragile, manual-intervention process into a reliable, repeatable one.
• Owned the architecture and stability of the critical financial data flows: multi-source valuation data feeding thousands of monthly client reports.
• Ran code reviews and provided technical mentorship across the team, and helped define engineering processes for the development lifecycle.
• Worked with geospatial property data at scale: PostGIS-backed queries, geographic boundaries, and map-based search (Mapbox).

A platform that reimburses fleet drivers for charging company electric vehicles at home. The system ingests vehicle telematics (charging sessions, battery and usage data) from the two leading fleet-telematics providers (Geotab and Samsara), normalizes it into a central data hub, and calculates accurate electricity costs per charging session, so drivers get paid quickly and employers get verifiable numbers.

• Built the ETL pipelines ingesting fleet-telematics data from the Geotab and Samsara APIs: designed, deployed and owned these services end to end.
• Used Python's asyncio throughout the ingestion layer: telematics APIs are I/O-bound and rate-limited, so concurrent async requests cut sync times dramatically compared to a sequential approach.
• Developed the core data API (FastAPI) serving normalized charging and vehicle data to the reimbursement engine, backed by PostgreSQL.
• Operated the services in production on Heroku, then handed ownership over to the client's in-house team at engagement end.

A mobile advertising platform delivering video ad campaigns for major consumer brands at massive scale, with a core delivery backend processing 800,000 requests per minute. The platform makes real-time decisions on which ad to serve, to whom, and under which campaign rules, where latency and reliability directly impact revenue.

• Optimized and maintained the mission-critical ad-delivery backend processing 800,000 requests per minute, where every millisecond of latency has direct revenue impact.
• Built advanced ad-targeting features, improving delivery precision and large-scale campaign performance.
• Worked under real high-traffic constraints: caching strategies, query performance and defensive design for systems that cannot go down during live campaigns.

A US software consultancy serving clients across industries: executive recruiting, sports training, e-commerce, and design. I worked as a senior engineer across multiple long-term engagements, owning projects end-to-end: scoping with the client, architecture, implementation, and delivery. Each engagement meant a different stack and a different business domain, from custom CRMs to desktop applications and systems audits.

• Designed a custom Ruby on Rails CRM for a high-profile executive-recruiting firm in New York.
• Built a Node.js / Electron desktop application for an online sports-training startup.
• Delivered technical consulting and a full systems audit for a leading Los Angeles design company, including complex Shopify integrations.
• Built and maintained Python/FastAPI backend services for multiple clients across multi-month engagements.

A test-preparation platform for the GMAT, serving students worldwide with practice exams, prep courses, and admissions consulting. The core product was a computer-adaptive practice exam engine designed to replicate the real GMAT experience, with question selection and scoring algorithms that adapt to each student's performance and closely match the official exam's behavior.

• Developed and maintained Ruby-based adaptive algorithms simulating the GMAT exam, with question selection and scoring logic closely matching the official test's behavior.
• Owned stability and new feature delivery across a diverse stack, including Ruby on Rails and PHP, for business-critical applications.
• Built hybrid mobile applications, bringing the exam-prep experience to mobile devices.