Tamara Orlich

Borah Studio is the internal operations platform I designed and built for my agency, Borah Digital Labs: a Laravel + Livewire + Filament application that handles project management, client reporting, income and expenses dashboard for management, quote generation assisted by AI, and legally compliant time tracking under Spanish labor law, including clock-in/out records, correction workflows with manager approvals, overtime detection against contracted hours, and PIN co-authorization for audit integrity. Everything is connected to Discord interactively via webhooks and custom apps.

I extended it with two native clients: Borah Spotlight is a keyboard-first desktop command bar (Tauri v2, Rust, React 19, TypeScript) summoned by a global hotkey: it creates tasks in a self-hosted Plane instance with token-based parsing (projects, assignees, priorities, natural-language dates) and drives the full time-tracking flow through slash commands, syncing with the Studio API via Laravel Sanctum. The mobile app, built with NativePHP (Laravel, Inertia, React), lets them manage each employee's tasks in Plane from iOS and Android.

I designed and built a multi-tenant SaaS platform that automates the daily operations of a Spanish customs brokerage handling import/export clearances. The system ingests email from different agencies' email accounts (Gmail API with OAuth and domain-wide delegation, plus IMAP fallback) and routes them to the right client and clearance file. Each file follows a typed document checklist with auto-completion, versioning with change-request history, and reusable templates per operation type.

Clients get their own portal to upload documents and track clearance status, and the platform integrates directly with AEAT (the Spanish tax agency) via SOAP web services secured with mutual TLS for electronic customs declarations. Built with Laravel, Inertia.js, React, and Tailwind CSS, covered by a Pest test suite.

Built a multitenant SaaS platform that lets catering companies automate their entire quoting workflow.

Companies can configure the quote through a guided wizard: the event date, event type, service style, menus, beverages, and staffing.

Then, the platform's pricing engine calculates a full budget in real time using different formulas based on product type, with all the logic customizable from an admin panel.

The admin panel is built with Filament, and the app uses Livewire for all reactivity. Quotes are rendered using Cloudflare Browser Rendering, so no headless Chrome is needed, and a corresponding Odoo quote is created. The platform is covered by Pest v4 unit, feature, and browser test suite.

I designed and built the complete web presence for Fusion Dance Training, a dance instruction and personal training business. Starting with just a logo, I created three distinct design directions for the client to choose from, then developed the selected concept into a fully produced site.

The site is built on Laravel with Tailwind CSS and a Filament admin panel, giving the owner full control over every piece of content: services, class schedules, testimonials, locations, and site settings, all without touching code. It's fully bilingual (Spanish/English) using Spatie Translatable, so each record is managed once and served in both languages.

I also implemented: WCAG 2.1 AA accessibility compliance, structured data (Schema.org) and SEO optimization, responsive design across all devices, a contact form with email notifications and submission tracking in the admin panel, and privacy-friendly analytics using Fathom Analytics.

I conducted comprehensive research on RAG evaluation frameworks to optimize a Spanish-language search system. The project involved analyzing four leading frameworks (RAGAS, LlamaIndex, UpTrain, DeepEval) across 12 metrics, including faithfulness, answer relevancy, and context relevancy.

Through systematic testing with production data, I discovered critical evaluation inconsistencies where frameworks provided contradictory scores for identical queries. My analysis revealed that 75% of performance issues originated from context retrieval rather than answer generation, providing crucial insights for system optimization. I documented the mathematical formulations and prompt structures underlying each metric, exposing the circular dependency problem of using LLMs to evaluate LLM outputs.

The deliverable included a comprehensive presentation explaining complex RAG concepts using accessible analogies, making technical content understandable for non-technical stakeholders. I quantified evaluation costs and established a monitoring pipeline for continuous system improvement. This research directly informed optimization strategies that improved retrieval accuracy and overall system performance.

I designed and developed a Python package that extracts and categorizes text content from PDF documents using ML classifiers and boundary detection algorithms. This solution optimizes RAG pipelines by addressing poor PDF text extraction that was degrading client system performance.

The package implements two random forest models: a section title detection model that formats titles as markdown, and a line-joining model that merges broken text lines. These ML components achieved a 75.2% overall error reduction.

Key technical achievements include automatic page rotation detection and correction, chronological text order preservation through PDF splitting, and improved format recognition for tables, figures, and links. The solution converts poorly structured PDF text into clean markdown that LLMs can process effectively, significantly improving retrieval accuracy in RAG systems.

I handled the complete development lifecycle, including algorithm design, model training, documentation, and PyPI distribution as an open-source package. The solution features robust error handling, configurable processing options, visual tools with colored boundary markers, and support for both single-file and batch directory processing.

Designed a multi-agent system in n8n that ensures any text aligns with the company's brand voice guidelines through orchestrated collaborative review. The system enables any person in a company to produce brand-consistent content (like emails, blog posts, and social media).
The architecture features six AI models: one orchestrator managing four specialized reviewers and one writer. I researched model capabilities using LMSYS Arena benchmarks to strategically assign roles, with each configured with fallback alternatives for API reliability.
The orchestrator manages the review process: the first reviewer analyzes the text, then shares their feedback with the next reviewer, who adds their own perspective, and so on until all reviewers have contributed. The orchestrator collects all feedback and sends these to the writer to revise the text. This process repeats until all reviewers approve the final version.
The n8n implementation features webhook integration for brand voice document retrieval, agent loop architecture, and structured JSON outputs, making it adaptable across industries.

I conducted a comprehensive benchmark study evaluating 11 reranking models for a Spanish-language RAG system, analyzing performance across 1,000 queries from the MS MARCO multilingual dataset. The project aimed to identify the optimal model balancing accuracy, speed, and multilingual capability for production deployment.

Through systematic testing, I discovered that jina-reranker-v2-base-multilingual achieved the highest accuracy with 91.1% of correct answers in top-5 chunks and an MRR of 0.3677, though at 2.6 seconds per query. I identified that the current production model (ms-marco-TinyBERT-L-2-v2) was underperforming due to a lack of Spanish training data, achieving only 56.4% accuracy despite a speed of 0.4 seconds.

I provided detailed recommendations balancing speed and accuracy requirements, suggesting jina-reranker-v2-base-multilingual for accuracy-critical applications and InfiniFlow/bce-reranker-base_v1 for balanced performance at 3.3 seconds with 80% accuracy. The evaluation framework I developed now serves as the standard for model selection in the client's ML pipeline.

I developed an intelligent email personalization system that transforms standard outreach templates into targeted, context-aware communications by automatically researching prospects and tailoring messaging to their specific needs and achievements.

The system leverages OpenAI's GPT-4.1 with web search capabilities to conduct real-time research on target companies, analyzing their recent achievements, digital presence, values, and potential pain points. Through a multi-stage prompt engineering pipeline, I created specialized agents that identify optimal service matches, select relevant case studies from our achievement database, and craft personalized value propositions that resonate with each prospect's specific situation.

The implementation features four interconnected prompting stages: company analysis and language detection, service-achievement matching based on prospect needs, highlight extraction focusing on collaboration opportunities, and final email composition with tone optimization. Each stage uses structured JSON schemas to ensure consistency and quality.

This automation streamlines the outreach process by eliminating manual research and ensuring consistent, personalized messaging across all prospect communications.

I developed a statistical framework to optimize RAG system efficiency by establishing thresholds for discarding irrelevant chunks before LLM processing. Using 1,000 questions (500 relevant, 500 irrelevant) across a Spanish document, I analyzed score distributions from Qdrant similarity search and multiple reranking models to identify optimal filtering thresholds.

My analysis revealed distinct score patterns between relevant and irrelevant content, with Qdrant showing relevant answers averaging 0.876 versus 0.768 for irrelevant passages. I established model-specific thresholds: 0.75 for Qdrant similarity search, 0.3 for ms-marco-TinyBERT, and 0.07 for jina-reranker-v2-base-multilingual, balancing precision and recall requirements.

A critical finding was that 80.56% of correct answers appeared in top-5 positions using similarity search alone, questioning the necessity of computationally expensive reranking for certain use cases. I also identified poor text chunking quality as a major problem, providing specific recommendations for performance improvements.

Over two years as a front-end developer, I have built and delivered web and mobile applications using modern frameworks, demonstrating versatility across different tech stacks and business domains. My portfolio spans social platforms, enterprise solutions, and consumer applications, all built with a focus on responsive design and user experience.

Projects included an Ionic Angular application integrating social features, real-time messaging, and multimedia content for traders; a micro-frontend customer portal using Angular, Nx, and NgRx for an insurance company; responsive websites with NuxtJS and TailwindCSS for music catalogs; mobile applications for the hospitality sector featuring real-time updates, personalized content, and appointment management; and an eco-friendly food delivery platform with QR code scanning and Stripe payment integration.

Throughout these projects, I consistently implemented responsive designs across web and mobile interfaces, integrated third-party services including payment gateways and real-time communication tools, and maintained clean, scalable codebases using modern development practices.