Jhovahn
Gibbs

Full-stack engineer. Applied AI.

Previously at Meta on Instagram Reels ads and Ads Manager. I work in React, Next.js, GraphQL, FastAPI, Rails and Kotlin, and I'm currently building applied AI tools.

San Francisco, CA Open to full-stack, forward-deployed & applied AI roles

Experience

2019 – 2025

Meta

Software Engineer

Oct 2024 – Oct 2025
  • Launched a mid-scene ad feature in Ads Manager for Instagram Reels, plus a matching timeline-advance feature on Instagram Android.
  • Built automated testing for Reels ads across ad creative types.
  • Ran Android UI experiments for Instagram ads in Kotlin (disappearing top gradient, gesture-exclusion zones).

TekSystems at Facebook

Software Engineer

Oct 2020 – Oct 2022
  • Built an internal React/GraphQL/Hack tool for adding to Business Products Engineering data pipelines.
  • Developed XHP components for the CMS behind the Facebook and Instagram for Business sites.
  • Shipped features for the Static Signal platform and Marketing Home internal tools; wrote regression test plans for CMS changes.

LTSE

Software Engineer

Jul 2019 – May 2020
  • Built an Excel-to-PostgreSQL ETL process with ActiveRecord.
  • Designed and deployed a record-cloning process in Ruby on Rails.
  • Delivered full-stack features from stakeholder requirements, test-driven.

Projects

Applied AI

Job Pipeline RAG

A Q&A system over a personal document set, built on FastAPI with local sentence-transformer embeddings and a cosine-similarity index written from scratch. Exact and numeric questions go to a structured data parser instead of embeddings.

A rerunnable eval harness took answer accuracy from 20% to 100% over two debugging passes. It has a 12-test pytest suite and runs in Docker.

Python · FastAPI · sentence-transformers · NumPy · pytest · Docker
100%Eval accuracy
12pytest tests
2Debug passes

Clarity

Voice transcription · Live

Record in the browser and get a verbatim Whisper transcript. A Claude pass then removes filler words, and you can toggle between Clean and Verbatim views, with removed words marked rather than lost.

The clean pass can only delete. Claude returns the exact spans to cut, and the server removes only text it finds verbatim, so no word is ever rewritten or added.

The repo also has its own coding agent. A GitHub Actions pipeline built on the Claude Agent SDK triages each issue and posts a plan. Once the plan is approved, it writes the change and opens a PR after CI passes. A person always does the merge.

React · Vite · TypeScript · Express · Whisper · Claude · Agent SDK · GitHub Actions
Live ↗ GitHub ↗

Morgenwort

German speaking practice · Live

A daily German word app. Each day serves a themed lesson plus spaced-repetition reviews. You say each example sentence aloud, and it scores what the browser's speech recognizer heard against the target sentence.

Claude phrases the feedback tips from human-written notes but never touches scoring. Progress is stored in the browser, so the server stays stateless.

React · TypeScript · Express · Web Speech API · Claude · Vitest
Live ↗ GitHub ↗

Skills

Languages

  • TypeScript
  • Python
  • JavaScript
  • Kotlin
  • Hack

Frameworks

  • React · Next.js
  • FastAPI
  • GraphQL
  • Express
  • Ruby on Rails
  • Android

Infra & AI

  • Claude API · Agent SDK
  • RAG
  • Whisper · Web Speech API
  • GitHub Actions
  • Docker · RabbitMQ

Data & testing

  • PostgreSQL · MySQL
  • pytest
  • Jest · Mocha · Chai
  • Git · Mercurial

Education

Hack ReactorAdvanced Software Immersive · San Francisco
Pace UniversityBBA, Finance

Get in touch

jhovahngibbs@gmail.com LinkedIn GitHub