Julián Antonucci

Primary backend engineer on a high-traffic survey and voting platform built for major US broadcast networks. During live broadcast events — the exact moment traffic peaks — votes were silently disappearing.

What broke

A legacy DynamoDB-to-MySQL migration had left a race condition in the vote-write path. Under the concurrent write volume of a live broadcast, writes overlapped and votes were lost with no error, no alert — just a count that didn't add up after the show aired.

What I did

• Traced the loss to the migration's overlapping writes and rebuilt the path around Redis-based vote buffering, collapsing concurrent increments into atomic operations before they ever reached MySQL.
• Optimized the critical queries in the read/aggregation path and configured CPU-based EC2 autoscaling to absorb live-broadcast traffic spikes.
• Validated the fix with JMeter load tests at 30,000 votes/minute — zero data loss, a 5x increase in processing capacity over the previous ceiling.
• Removed the dead code left over from the original migration once the new path proved out in production, closing off a source of latent bugs.
• Built and extended the platform's social distribution integrations (Instagram, TikTok, Facebook, Vimeo) for direct video/image publishing from its media management platform.

A Node.js/TypeScript product that turns uploaded images into analysis reports, built with two partners as a personal project outside my day-to-day work. I lead the backend and I'm moving the whole platform off one droplet onto AWS serverless while it keeps serving traffic.

Where it started

A single DigitalOcean droplet ran the API, the Postgres database and image processing together. Deploys were a git push that SSH'd into the box — no CI gate, no rollback, no preview. Every upload processed synchronously, so traffic spikes queued behind image work and the API slowed for everyone.

What I did

• Introduced a StoragePort interface with an S3Storage implementation, moving image and preview bytes out of Postgres and onto S3. Read paths now serve signed URLs instead of base64 payloads.
• Added previewKey / imageKey columns plus a backfill, and relaxed the legacy base64 NOT NULL constraint so old and new rows coexist through the cutover.
• Redesigned the analysis pipeline around a Queue port with a pluggable backend — one Postgres implementation today (LISTEN/NOTIFY + FOR UPDATE SKIP LOCKED), one SQS implementation for AWS, same worker code either way.
• Designed the target AWS architecture end to end in CDK: App Runner for the API, the existing worker Docker image redeployed as a container-image Lambda behind SQS, EventBridge routing job-completion events to SES for email notifications, and Neon (serverless Postgres) in place of a managed instance.
• Killed a class of "stale app after deploy" bugs — a cached index.html and orphaned service workers that kept serving a months-old bundle to returning users.

Architecture decisions

• App Runner over API Gateway + Lambda for the API — the request path is rarely idle at this traffic, so a container that autoscales on its own beats paying Lambda's cold-start tax on every request.
• Neon over RDS — at this volume an always-on managed instance would cost more than the workload itself; Neon bills for compute actually used and scales to zero between requests.
• No VPC, no NAT gateway — App Runner, the worker Lambda and Neon all talk over public TLS endpoints. A private database would have forced a NAT gateway that alone costs more than the database.
• Worker ships as the same container image the droplet ran, redeployed to Lambda — no parallel build pipeline to maintain for two runtimes.

Co-founded a software studio serving the graphic arts industry and designed its flagship product: a print optimization system that lays out jobs to minimize material waste.

What I did

• Designed a microservices architecture: a Node.js/TypeScript backend orchestrating async PDF generation jobs over RabbitMQ, and a Python microservice using Ghostscript to compute layouts that minimize material waste.
• Built real-time order-status updates over WebSockets and moved generated artifacts to S3.
• Cut PDF data-extraction time 75% with a Node.js REST API backed by the Python microservice.
• Cut internal approval-workflow completion time 50% by replacing polling with real-time WebSocket (Socket.IO) file flows.
• Cut local environment setup time 40% by containerizing the stack with Docker and managing processes with PM2.