I build AI agents and predictive models, and I start every one by asking what it costs when it's wrong.
- BASED
- Atlanta, GA
- NOW
- Building Vuee for Okearin Mart, Lagos
- LOCAL
- Atlanta (Eastern Time)
Introduction
I'm Jovi, an AI and ML engineer in Atlanta. I studied actuarial science, interned in credit risk in London, and spent two years as a data analyst on 1,200 telecom mast sites. Now I'm building Vuee, an AI tool for a Lagos distributor that waits for a manager's sign-off. I also built Fifi, and I shoot short films.
Selected work
Three cuts I'd show you first.
- Vuee
- Fifi Shipped an in-browser MVP, pitched it at iLaunch, rebuilt the core as an agent-run research pipeline, and paused it at the limit of the data available.
- Predictive maintenance at Makasa Sun Fault flags at 87% precision; weekly reporting cut from 40 hours to 10.
- Teaching Claude Code to cut like me A 30-second vertical cut from seven studio clips, planned shot by shot. Now extended to event recaps with dialogue.
( REEL 2 — SELECTS )
Three cuts I'd show you first.
Open the binMore projects
-
Echo Claw
An open-source, Jarvis-style voice assistant that runs on my own hardware: OpenClaw and Claude for the thinking, Home Assistant for the house, an Echo Show for the screen.
PHASE 1 OF 6- OpenClaw
- Home Assistant
- Claude
- Whisper
-
Fantasy Football MCP
The data backbone for a fantasy-football chatbot: live NFL stats, injury reports and roster memory, exposed to the agent as MCP tools. I learned MCP from zero in a week and debugged SSE versus Streamable HTTP in production.
CLASS PROJECT- FastMCP
- Python
- Sleeper API
- PostgreSQL
-
Bar-hopping route optimizer
Tell it "yo let's hit Chimy's and Cricket's at 9" and it plans the night: a small language model fine-tuned to understand casual requests, plus a simulation that orders the bars to cut your wait.
CO-BUILT- Phi-3.5-mini (3.8B)
- LoRA / PEFT
- PyTorch
- FastAPI
-
Insurance lifetime value
Customer lifetime value, applied to insurance risk. In that data, visit frequency predicted total cost better than clinical risk scores, and regional differences outweighed plan types.
CAPSTONE · TEAM OF 3- Python
- Canonical correlation
- Gaussian mixtures
- PCA
Experience
EDIT DECISION LIST — 9 EVENTS
-
001 VUEE V2 NOW Founder & engineer — Vuee, for Okearin Mart (Lagos) · Atlanta · remote
- Vuee is an AI decision platform for Okearin Mart, an import-wholesale grocery distributor in Lagos.
- Every action the system proposes, such as a transfer invoice between warehouses, waits for a manager's sign-off before it runs.
-
002 TTU V1 2025-09 2026-05 Graduate Assistant — Texas Tech University · Lubbock, TX
-
003 TTU A1 2026-05 MS in Data Science · GPA 4.0 — Texas Tech University, Rawls College of Business · Lubbock, TX
- Applied Multivariate Analysis · Time Series Analysis · Operations Research · Database Management Systems
-
004 FIFI V2 2025 2026-10 Founder & lead ML developer — Fifi · Lubbock → Atlanta
- Fifi started as a virtual fitting room: a few inputs, a body-shape estimate, and a 3D model rendered in the browser.
- Eight Claude Code subagents with distinct roles research the biology: an endocrinologist, an epidemiologist, a source verifier that checks every DOI via CrossRef, an integration analyst, an adversarial reviewer and three specialists.
- Seven validation hooks enforce the pipeline's rules in code, exiting 1 to block an experiment or 2 to flag it.
-
005 MAKASA V1 2023-06 2025-05 Data Analyst — Makasa Sun · solar and hybrid power for telecom mast sites · Remote
- A generator fault model (ARIMA and XGBoost) flagged sites due for an engineer check-up from fault-code history at 87% precision.
- Weekly reporting for 1,200+ telecom mast sites moved from manual Excel to Python scripts and automated pipelines (SQL, AWS S3, Redshift), cutting weekly reporting time from 40 hours to 10.
- Analyzed 50K+ monthly site payment transactions in Python, improving forecast accuracy by 10% and reducing revenue variance by $120K a year.
- Led the predictive maintenance project end to end, agreeing deliverables and timing with leadership, and managed a 3-person finance team.
-
006 ESSEX A1 2023-07 BSc Actuarial Science — University of Essex · Colchester, UK
-
007 GTBANK V1 2022-06 2022-08 Risk Analysis Intern — Guaranty Trust Bank (UK) · Credit Risk, Retail Lending · London
- Cleaned and validated 500K+ transaction records in SQL, cutting data preparation time by 60%.
- Built retail loan default models in R across 10K+ borrowers (82% accuracy), comparing logistic regression with decision trees and choosing on explainability, not accuracy alone.
- Prepared the loan default modelling dataset, joining customer, loan and repayment tables and setting the definition of default used by the model.
-
008 METRO V1 2021-06 2021-09 Customer Representative — Metro Bank (UK) · UK
- Account openings, closures and transactions; outbound calls promoting insurance products.
-
009 FILM A2 2026 NOW Videographer & editor — Freelance · Atlanta · Atlanta
- Portraits, birthdays and event recaps; I shoot, cut and grade.
Film
I shoot and cut short films in Atlanta: portraits, birthdays, events. The instincts are the same ones I use in engineering. Find the strongest material, cut on the beat, and don't let anyone see the seams.
Cut 04 How I taught Claude Code to cut like meAbout
I'm Jovi Ukwade, an AI and ML engineer in Atlanta.
I studied actuarial science at the University of Essex. In the summer of 2022 I interned in credit risk at Guaranty Trust Bank in London, building loan default models in R.
After graduating I spent two years as a data analyst at Makasa Sun, working with data from more than 1,200 telecom mast sites. I built a model that used fault-code history to flag sites due for an engineer's check-up, and I moved weekly reporting off Excel and onto Python and automated pipelines. That took it from 40 hours a week to 10.
Then I did a master's in data science at Texas Tech, worked as a graduate assistant there, and finished with a 4.0.
Now I'm building Vuee with Okearin Mart, a wholesale grocery distributor in Lagos. It answers questions about stock and sales and proposes what to do next, and nothing goes through until a manager signs off.
I also shoot and edit short films around Atlanta: portraits, birthdays and events.
Things I believe
- Guardrails only count if code enforces them.
- An agent can do everything statistically right and still be completely wrong.
- Before you ship a model, know who pays when it's wrong, and how much.
- Agents don't fill in blanks the way people do. Specify, or they'll guess.