Josue – Golang, Python, Kubernetes

A healthtech platform for primary-care practices, where front-desk staff and providers lose hours every day to phone calls, scheduling and manual quality reporting. AI agents handle incoming calls, appointment booking, reminders and patient follow-ups. On top of that sits a quality-improvement module for EMR-based clinics. It analyzes clinical indicators with LLMs and a RAG pipeline, flags cases for clinician review and backs each one with source-cited evidence. The platform is privacy-first: patient data is de-identified before any AI touches it, AI access is read-only, and every action needs human approval.

• Built and launched an agentic quality-improvement platform for OSCAR EMR clinics end-to-end (LLM-driven indicator analysis plus a RAG pipeline), delivering monthly QI charts with source-cited provenance for every flagged case.
• Cut infrastructure costs by 60% by right-sizing capacity with autoscaling and moving background workloads to spot instances.
• Designed the privacy-by-design architecture: a de-identified semantic layer, human approval gating and read-only AI access to patient data.
• Built the agentic automation for call handling, appointment booking, reminders and follow-ups.
• Own the technical strategy as a hands-on fractional CTO, from architecture to delivery.

A real-time fraud detection SaaS for online operators who lose revenue when users create multiple accounts to exploit free trials and sign-up bonuses. Each sign-up is scored across 150+ signals, including device fingerprinting, network intelligence, behavioral patterns and consortium data shared across operators. The platform returns a single fraud verdict through an API, fast enough to run inline in sign-up and checkout flows without slowing them down.

• Designed and built the platform end-to-end as the sole engineer: a 150+ signal detection engine that returns a fraud verdict in 67ms median with 94% accuracy.
• Chose in-process Go fan-out with bounded worker pools and Redis over chained gRPC microservices to meet a ~60ms p99 latency target, trading strict consistency for speed.
• Built fail-open handling for third-party signal timeouts, so the platform hit its latency targets without causing any sign-up or checkout drop-off.
• Implemented the detection signals: device fingerprinting, network intelligence, behavioral analysis and cross-operator consortium data.
• Set up Stripe billing and deployed the platform on GCP with Kubernetes and Docker.

An AI training and evaluation program in which experienced engineers stress-test large language models on real-world Go development. Engineers review model-generated code, design multi-step tasks for AI agents and write hard algorithmic challenges that expose weaknesses in AI code generation. The resulting evaluations and datasets are used to improve how models write and reason about code.

• Evaluated high-complexity, LLM-generated Go code for concurrency correctness, memory efficiency and production reliability under load.
• Designed and validated AI agent workflows for multi-step web-browsing tasks, making them more robust in high-variability scenarios.
• Engineered a feedback loop that improved the reliability of AI code generation and reduced error rates in automated developer tooling.
• Authored competitive-programming-style challenges (LeetCode/Codeforces level), each with a full test suite and an optimal reference solution.

A subscription micro-SaaS for real estate agents who spend a lot of time writing property listings by hand. Agents upload photos of a property, and multimodal AI analyzes the images in real time and generates ready-to-publish listing copy.

• Built and launched the product end-to-end in Go and HTMX as a cloud-native app on GCP and Kubernetes.
• Integrated a multimodal LLM (Gemini) to turn property photos into listing copy in real time.
• Implemented S3-backed image storage and Stripe subscription billing.

A sports media platform serving millions of users across TV broadcast, digital products and betting partners. Its core stats system ingests live-game events and distributes processed statistics in real time to internal clients such as broadcast graphics panels and betting and e-gaming partners. The platform also runs the user data services behind its consumer apps.

• Built the Go microservices that process live-game stats and deliver them in real time to TV broadcast panels and betting/e-gaming partners, as one of a two-person core stats team (Kafka self-hosted on AWS, PostgreSQL, EKS).
• Designed Kafka-based dispatch of typed live-game events to consumers with different data needs.
• Designed and delivered a high-availability user data microservice in Go, backed by Cassandra and Redis, for millions of users.
• Added redundancy and failover patterns across the microservices to keep them up under peak game-day load.
• Turned business requirements into technical specs with cross-functional teams and mentored junior engineers on distributed systems.

An internal HR application for a consulting firm whose workforce is organized into squads, chapters and guilds. It matches mentors with mentees based on skills and preferences, lets employees find people quickly across teams, and centralizes employee profiles, CVs and photos.

• Designed and built the mentor–mentee matching algorithm, pairing people by skills and preferences.
• Built a search API across squads, chapters, guilds and teams, and added Redis caching for frequent queries to speed up search.
• Developed ETL tasks that pull data from internal applications into the HR app.
• Implemented storage of CVs and profile pictures in Google Cloud Storage.

Internal platforms that run a European cloud provider’s data centers. They poll and monitor tens of thousands of physical devices, measure energy efficiency (PUE) and bill customers for the resources they consume.

• Re-architected billing into a distributed, event-driven system, cutting processing time from 3 days to 1 day (a 66% reduction).
• Scaled the device polling platform to 30,000+ devices with Celery and Redis while keeping error rates under 1%.
• Built Go tooling for data center energy monitoring: a Modbus RTU-to-TCP converter and a PUE calculator API.