Bright Uzosike
Data Scientist · ML EngineerManchester, UKhello@brightuzosike.devbrightuzosike.devlinkedin.com/in/bright-uzosikegithub.com/michizler
SUMMARY
Data scientist who builds, ships and maintains explainable ML systems that add business value. I evaluate before I trust, so gold sets, cross-validation and a written list of limitations ship with every model. I hold an MSc in Data Science with Distinction from the University of Salford and have six years of experience across analytics, BI and machine learning.
EXPERIENCE
Research Data Scientist · Rebel Hive
May 2026 to present- Sole data scientist on a live client project, taking an ambiguous brief to an NLP diagnostic tool now in commercial use by the client.
- Engineered a 76-feature matrix from 38 regulatory and corporate documents. Classifiers predicting Ofsted outcomes from language alone reached macro-F1 of 0.82 to 0.91 under 5-fold stratified CV on 24 inspection reports.
- Validated with pre-specified hypotheses, non-parametric tests, robustness checks and an inter-rater reliability harness, and reported counter-theoretical findings rather than overclaiming.
- Shipped a single-command pipeline with a pinned environment and fixed seeds, served through FastAPI and Streamlit, with a replication guide and a forward data contract.
AI Data Specialist (RLHF & SFT) · Mercor
Aug 2026 to present- Score and rank 2 to 8 candidate model outputs per prompt against multi-axis rubrics, writing auditable rationales that feed reward-model training.
- Flag reward-hacking failures where polish hides substance, such as fabricated tables, unsupported claims and arithmetic errors.
- Build Azure Data Factory pipelines for streaming recommendation systems, and Informatica DEI mappings, to per-task SLAs.
AI Training Contributor (SFT & RLHF) · Outlier
Sep 2025 to present- Passed the Generalist SFT and RLHF screening. Evaluate model outputs for accuracy, instruction-following and safety, writing structured feedback used in retraining.
Data Science Consultant (internship) · Amdari
Jan 2025 to Jun 2026- Delivered end-to-end ML engagements for several clients, covering scoping, EDA, feature engineering, modelling, evaluation (precision, recall, F1, AUC-ROC) and stakeholder reporting.
- Built ETL/ELT pipelines in Azure Data Factory, and agentic AI and RAG automations on Microsoft Power Platform.
- Turned model outputs into dashboards and presentations for non-technical stakeholders. Retained beyond the initial term to deliver for two further clients.
GIS Analyst · Nationhood Podcast
Mar 2025 to Apr 2025- Georeferenced and digitised map files (GeoTIFF, GeoPackage) in QGIS and prepared them for web embedding.
HR Data Analyst to Branch Director · AIESEC Nigeria & UK
Apr 2022 to Jul 2025- Rose from Talent Manager and HR Data Analyst to Branch Director, leading distributed teams of 50+ members, then led the HR and organisational development team at AIESEC UK.
- As Branch Director, led AIESEC in Enugu to the top spot in operations out of 13 branches in Nigeria in July 2024.
- Built KPI dashboards for all 13 national branches and raised data consistency by 10% by linking reporting to cleaned source data.
- As HR Team Lead, raised data integrity to 93% through validation and review of the team's datasets.
Business Intelligence Analyst · TRACKER.NG
Jan 2023 to Aug 2023- Automated Tableau dashboards joining sales, customer service and operations data, cutting a decision-making bottleneck from 35% to 15%.
- Added validation logic that raised data integrity by 30% and halved ad hoc correction requests.
- Liaised between business development, operations and IT, cutting project delivery timelines by 20%.
Software Engineering Intern · ALX Africa
May 2022 to Jul 2023- Completed a 12-month software engineering programme, building full-stack Python and JavaScript applications on Linux.
Data Analyst Intern · Trainiton Analytics
Oct 2020 to Apr 2022- Automated repetitive data processing in Python, and co-led analyst training with 85% retention over the final six months.
SELECTED PROJECTS
AlloyTower AVM
8.0% MAPE with an assessment, 44.9% without (5-fold CV). Every estimate is explained with SHAP and served by a five-endpoint FastAPI service in Docker that redeploys on every push.
case study →VoyageRail fare desk
Fare model that quotes from what a booking desk knows at quote time and explains every price with SHAP. Gradient boosting reaches R² 0.90 against a 0.71 linear baseline.
case study →LLM Feedback Analyzer
Schema-validated Claude output with retries, backoff and a safe fallback. 13 of 15 correct on a hand-labelled gold set for sentiment and churn risk.
case study →US Bancorp volatility
Holt-Winters and ARIMA failed Ljung-Box. ARIMA(1,0,1) with sGARCH(1,1) passed (p = 0.2369) and gave a 365-day price and volatility forecast.
code ↗