Juan – Python, LLM, AI agent development

A multi-tenant medical-practice management system for a multi-clinic gynecology practice, designed and built independently end to end and still in active development. It replaces Word-document patient records with structured visit notes, adds per-clinic billing, and integrates appointment scheduling with Google Calendar. The system was designed multi-tenant from the first schema so additional clinics can onboard without code changes, with all AWS infrastructure defined in Terraform ahead of launch.

• Designed the full product independently: requirements gathered from practicing physicians, data model, API, UI, and infrastructure.
• Built the backend with Django and PostgreSQL using a multi-tenant schema with per-clinic data isolation and billing.
• Built the frontend in React, covering structured visit notes, patient history, and scheduling workflows.
• Integrated Google Calendar for appointment scheduling.
• Defined all AWS infrastructure as code with Terraform ahead of launch.

The internal AI agent platform of a US construction-technology marketplace, serving the sales, operations, and support teams. The platform provides a generalizable agent service and an agent harness with guardrails that runs arbitrary numbers of agents concurrently, plus a unified communications layer that consolidated customer conversations across channels into a single source of truth. It is the foundation the production automations across the business are built on, replacing manual, multi-day workflows with agent-driven ones.

• Designed and implemented the architecture for a generalizable agent service platform using a FastAPI service deployed on Kubernetes.
• Implemented agent harness integrations and concurrency locks with autoscaling to support arbitrary numbers of agents running simultaneously.
• Designed and implemented a heuristics-plus-LLM automation for user-reported bugs and Sentry triage through an event-driven architecture using Amazon SQS, Amazon SNS, and a multi-step AWS Lambda in Python.
• Built an interactive multi-step coding assistant with prompt templates for reliable generation of Django code using the OpenAI SDK in Python.
• Architected a unified communications platform that consolidated customer communications into a single source of truth with dual-consumer event flows.
• Re-architected the company’s Golang event bus to introduce true worker concurrency and graceful shutdown behavior.
• Led a team of 6 software engineers across cloud infrastructure, DevOps, security, platform, data, and AI initiatives.
• Led incident response management, enforced an incident response plan, and trained engineers through scenario simulations.

An AI-powered clinical decision-support system that won 1st place at the Saptiva AI Hackathon (Mexico Tech Week 2025). As a doctor completes a medical record during a consultation, the system retrieves official treatment guidelines and medication regulations in real time and surfaces suggestions and evaluations inline, grounding every recommendation in cited source documents rather than free model output.

• Built a FastAPI backend powering retrieval-augmented generation over a Weaviate vector database of official treatment guidelines and medication regulations.
• Designed the retrieval pipeline: embedding model selection, chunking tuned from 512 to 1024 tokens with 20% overlap after measuring recall and precision of top-k results, and metadata filtering to restrict search to relevant document classes.
• Built the Next.js frontend delivering real-time suggestions while the doctor completes the medical record.
• Won 1st place against competing teams within the hackathon timebox.

The company-wide data platform of a US construction-technology marketplace, serving finance, operations, and leadership. The platform replicated the transactional PostgreSQL databases into BigQuery through a custom CDC pipeline, transformed raw data into curated analytics datasets, and powered a 1,500-metric executive dashboard refreshed every 15-60 minutes. It also included a Monte Carlo simulation of the sales-and-operations funnel that let finance and leadership test acquisition-spend and hiring scenarios before committing budget.

• Architected a company-wide CDC replication platform using PostgreSQL audit triggers, watermark-based incremental extraction, GCS staging, and idempotent partition-bounded MERGE loading into BigQuery, with a schema-evolution-resilient JSON staging layer — 15-minute freshness SLA, operated 4+ years with near-zero maintenance.
• Migrated orchestration to Apache Airflow on AWS with separate staging and production, high-concurrency Lambda functions, and DB read replicas, modernizing a 1,500-metric pipeline to 30-60 min freshness with automatic recovery.
• Reduced BigQuery spend by $35k per year through time-partitioning, clustering, partition-bounded MERGE operations, pre-aggregated gold tables, and per-query cost attribution with monthly audits.
• Led the deprecation of a legacy warehouse using a full consumer-dependency inventory, dual-run reconciliation, and a staged cutover with rollback paths — zero consumer-reported outages.
• Built a Monte Carlo simulation of the sales funnel in NumPy, modeling Poisson lead arrivals, Bernoulli conversions, and Weibull waiting times fitted from warehouse data, used for budget and capacity planning.
• Led the data team and doubled its size, hiring and ramping engineers and analysts through structured interviews and onboarding bootcamps; created the data-analyst career ladder used for objective performance reviews.

The reporting and analytics layer of a US construction-technology marketplace, used by the operations and finance teams and by company leadership. The work replaced manual, spreadsheet-based reporting with scheduled, reliable data workflows, and consolidated over 1,500 regional sales, operations, and finance metrics into a single executive KPI dashboard used for weekly leadership business reviews.

• Engineered ETL pipelines and automation scripts using Python, SQL, Google Apps Script, and BigQuery, replacing manual spreadsheet reporting for the operations and finance teams.
• Developed and maintained the company-wide executive KPI dashboard on Google Data Studio over curated BigQuery datasets, consolidating 1,500+ metrics into a single source of truth.
• Established software-engineering standards for analytics by migrating all scheduled queries into a version-controlled GitHub repository with pull-request reviews and repo-based deploys — still the team standard 4+ years later.
• Began leading technical interviews and mentoring junior analysts.