Daniel Fonnegra Garcia

Developed a highly customizable CMS with multitenancy and white labeling that Metrofibre uses to manage land use agreements. I used Flask to expose the REST API, SQLAlchemy, and PostgreSQL for storage management, and Vue and Vuetify for the front end.

An Android app to register animals found during inspections, post pictures, and generate reports periodically in Microsoft Word and Excel and the project was implemented with Python, FastAPI, SQLAlchemy, and Flutter and deployed in an Amazon EC2 instance with the AWS CDK.

Mineros SA needed an autonomous solution to generate dredge extraction maps and calculate extraction volumes from operator reports.

I architected and developed a serverless website within the AWS ecosystem. The user interface is built on the Vue framework and leveraged Amazon CloudFront for deployment. Python forms the core of the back end, interacting with a DynamoDB database, and is deployed using AWS Lambda with API Gateway for RESTful API services. Amazon Cognito provides secure identity management.

This deployment via AWS resulted in an extremely cost-efficient and maintainable web application that can incorporate future projects on the same platform.

The project involved automating multiple mean reversion strategies in Python where I crafted a high-performance asynchronous application adhering to clean code practices, ensuring easy addition of new strategies for clients. This was achieved using asyncio, HTTPX, WebSockets, and telnetlib3, integrated with services like Polygon.io, Flash SEC, and DAS Trader.

A client required assistance configuring a load balancer and ingress rules for a bare metal Kubernetes cluster as part of their NFT website initiative. Collaborating with their team, we established a two-node cluster using MetalLB and ingress-nginx, which effectively manages the load distribution via a round-robin strategy.

A bitrix24 application that connects an open channel to a WhatsApp line. This application is currently running in an EC2 t3.medium instance and can handle a throughput of 1,000 messages per minute.

The app orchestrates AWS Lambda, AWS SQS, and the EC2 instance to avoid message loss and increase availability.

A FastAPI web application deployed in AWS App Runner was down. The project had two goals:
• Bring the app back up.
• Onboard a new tenant to the platform and document it.

The application was successfully fixed, and a deep dive was required to figure out how to onboard the users.

I implemented an AI agent that explores datasets to find answers to user questions written in natural language. The results can contain GeoJSON columns that we will display on a map using tiling packages. We implemented the API in Python with FastAPI and PostGIS connected to Claude Haiku 4.5.

The application is hosted on AWS with various tenants. The main issue is with the back-end code, which is currently failing to communicate with the database. The client has access to the GitHub repository and database for troubleshooting.

I served as a Consultant and Tech Lead on Nuvant’s commercial demand forecasting platform for manufacturing decision-making. In Phase 1, I helped deliver an ARIMA forecasting baseline. In Phase 2, I focused on productizing the commercial forecasting model by building API endpoints, historical and forecast visualizations, and forecast editing capabilities. I inherited a 3-person engineering team that had not shipped a deliverable in six months and established daily production deployments with a two-week release cadence. The stack included TypeScript and AngularJS on the front end, Python with ARIMA on the forecasting layer, REST APIs, Kubernetes on GCP, and CI/CD pipelines.

I built Cortex, a data-agnostic capture system across Downloads and Gmail with natural-language Q&A as the sole retrieval surface. The taxonomy system learned types, contexts, and entities entirely from triage feedback with zero pre-seeding. I implemented two-stage classification with a relevance gate, label classifier, rule predicates, prefiltering, a 20-rule cap, and cascading taxonomy operations, including rename, merge, split, and deprecate workflows.

I built a Mac launchd daemon using Bun and TypeScript to run read-only Downloads and Gmail collectors, while a Vercel-hosted Next.js 16 web app served as the DB access boundary. Used Neon serverless PostgreSQL with pgvector (HNSW, halfvec 512-dim embeddings) via Prisma 7. I also implemented a client-driven two-phase ingest flow with reserve, upload, and confirm stages using Drive resumable uploads.

Heavy LLM work, including classification and chunking, was offloaded to Claude Code CLI on the Max subscription, while Anthropic Haiku handled Q&A synthesis. I integrated Clerk for authentication, and Langfuse tracing across classify, chunk, embed, and ask calls. I shipped seven phases in three days while keeping total runtime costs around $5–10 per month at personal usage volume.