John Noriega / Engineering case studies

Behind the systems.

What I built, why the boundaries matter, and where to inspect the work.

01 / Sales platform

From conversation to qualified lead.

$2M+ in sales supported since launch.

I designed and launched Billboard Source’s AI sales intake and call-routing platform, connecting inbound conversations, transcription, lead qualification, and Nutshell CRM data entry.

My role and the outcome

I began consulting in January 2025 and moved to full-time software engineering in July 2026. The product launched in late January 2026. The $2M+ figure is the cumulative platform sales outcome through November 13, 2026, spanning both roles—not a separate result attributed to the voice agent or a claim of sole credit for the sales.

Implementation and decisions

Next.js and TypeScript provide the application layer; PostgreSQL and Drizzle hold structured records. Twilio connects phone workflows, OpenAI APIs support extraction from conversations, and Nutshell integrates the resulting leads into the sales workflow. Vercel hosts the application.

The integrations connect distinct systems rather than hiding everything inside a model conversation. Structured records make the output useful to downstream sales operations, while explicit integration boundaries expose failures that need diagnosis.

Verification

I added GitHub Actions checks for linting, type checks, unit tests, and Playwright browser tests backed by disposable PostgreSQL. Retained failure artifacts support debugging; flaky-test rejection prevents a retry from concealing an unstable test.

Production business application. No public demo is offered here; the repository documents the implementation without exposing customer workflows.

02 / Python voice agent

Useful calls. Bounded failure.

I built a Python conversational voice service using Pipecat, OpenAI, and Twilio Media Streams to capture caller requirements and create Nutshell leads. Parameterized PostgreSQL queries retrieve location data.

Security and operational tradeoffs

I hardened webhooks and WebSocket sessions with Twilio signature validation, per-process concurrency limits, call deadlines, and overflow voicemail. Calls consume finite resources: bounding their lifetime and providing a fallback matters as much as the successful conversation.

Per-process limits are deliberately not a distributed quota. Scaling to multiple workers requires accounting for that boundary rather than assuming one process’s counter caps the entire deployment.

Delivery and verification

Non-root Docker containers, health checks, and an Nginx reverse proxy support deployment. OpenTelemetry and Jaeger tracing help diagnose call behavior; automated tests cover rejection and timeout paths.

No public call demo or separate revenue metric is claimed. Production phone numbers and telephony webhooks are not demo endpoints.

03 / This portfolio

A small, complete publishing system.

This site pairs server-delivered HTML and CSS with a Node.js/Express service, SQLite-backed field notes, and WebSocket support. The homepage is the live demo.

Decisions

Core career and project content is available without client-side rendering. Field notes load through a JSON API; administrative publishing uses authenticated endpoints. SQLite keeps persistence simple for a single-service deployment, with a corresponding need to preserve the database across releases.

Verification

Tests exercise real HTTP requests: public reads, authentication failures, authorized publishing and deletion, visitor metrics, and resume uploads. Security headers, request limits, and signed expiring admin sessions protect the publishing surface.