Skip to content
View michizler's full-sized avatar

Block or report michizler

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
michizler/README.md
Bright Uzosike, Data Scientist and ML Engineer. I build, ship and maintain explainable ML systems that add business value.

I'm a data scientist and ML engineer in Manchester. I build models that businesses can price, plan and make decisions with. I test them on data they haven't seen, explain every prediction, and ship them as services other teams can call.

What I build: Pricing and risk models (dynamic pricing, property valuation, customer churn), served behind typed APIs, containerised and deployed. Data platforms that account for every record, for NHS prescribing and health records.

Now: Research Data Scientist at Rebel Hive (sole data scientist on a client NLP tool) and AI Data Specialist (RLHF & SFT) at Mercor, reviewing model outputs for frontier training.

Background: MSc Data Science with Distinction, University of Salford (2026) · BEng Electronic Engineering, Second Class Upper.

Looking for: Data science or ML engineering roles in the UK.

Featured systems

AlloyTower AVM: values a property through a live API and explains every estimate with SHAP. 8.0% MAPE with an assessment, 44.9% without (5-fold CV). LightGBM, SHAP, FastAPI, Docker. FHIR to OMOP: turns health records into research data, then trains across trusts without sharing patients. AUROC 0.676 federated against 0.650 for one trust alone, across five simulated trusts on synthetic patients. dbt, Postgres, OMOP, Flower. NHS prescribing lakehouse: loads every NHS England prescription item and proves it matches the publisher. 216 million rows loaded over 12 months, and £10.35bn agrees with the publisher. Data Factory, Databricks, Delta, Power BI. VoyageRail pricing: quotes rail fares from what a clerk knows at quote time and explains every price. R squared 0.90 on holdout with gradient boosting, against a 0.71 linear baseline. scikit-learn, SHAP, FastAPI, Docker Compose. Rebel Hive diagnostic, a client NLP tool in use: measures the gap between how colleges describe themselves and how Ofsted describes them. Macro-F1 0.82, 5-fold stratified CV on 24 Ofsted inspection reports. spaCy, scikit-learn, SHAP, Claude API. LLM Feedback Analyzer: turns customer feedback into schema-validated JSON, and falls back safely when the model fails. 13 of 15 correct on a 15-review gold set, for sentiment and for churn risk. Claude Haiku 4.5, Pydantic, Python.

Try AlloyTower live with the app or the API docs. The free-tier API can take about 40 seconds to wake.
Try FHIR to OMOP live on the FHIR to OMOP Desk.
Rebel Hive is client work, so its code is private.

How I build

  • Evaluate before you trust. A hand-labelled gold set for the LLM analyzer, 5-fold CV for AlloyTower, and Ljung-Box residual tests that ruled out two of three US Bancorp models.
  • Explain every prediction. Each AlloyTower valuation comes back with the SHAP contributions behind it.
  • Document the limitations. AlloyTower's data-quality report ranks 11 findings by severity, including coordinates that turned out to be random.
  • Ship it. A live app and API that redeploy on every push, and services that start with one Docker Compose command.
How AlloyTower fits together
flowchart LR
  subgraph Training
    T["train_avm.py"] --> ML[("MLflow runs")]
    T --> B[("Model bundles A and B")]
  end
  U(["User"]) --> S["Streamlit app<br/>Streamlit Cloud"]
  S -- "JSON over HTTPS" --> A["FastAPI service<br/>Docker on Render"]
  B --> A
  A --> R{"Assessed value<br/>given?"}
  R -- yes --> MA["Model A<br/>assessment-aware"]
  R -- no --> MB["Model B<br/>fundamentals-only"]
  MA --> X["SHAP TreeExplainer"]
  MB --> X
  X --> O["Estimate, range<br/>and top drivers"]
  A --> C["/comparables<br/>similar properties"]
Loading

Toolkit

Languages: Python, SQL, R, JavaScript, C, Bash Software engineering: FastAPI, Flask, Pydantic, SQLAlchemy, React, Next.js, Vite, HTML, CSS, Streamlit, Git, GitHub, Linux, REST API design Data science & ML: pandas, NumPy, scikit-learn, statsmodels, SciPy, XGBoost, LightGBM, SHAP, spaCy, Hugging Face, Jupyter, tidyverse, caret, forecast, rugarch LLMs & evaluation: Claude API, Structured outputs, Retries and backoff, RAG, Agentic workflows, Gold-set evaluation, RLHF & SFT Data engineering: Azure Data Factory, Informatica DEI, Databricks, Spark & PySpark, Microsoft Fabric, BigQuery, SQL Server, PostgreSQL, MySQL, ETL/ELT MLOps & cloud: Docker, Docker Compose, Kubernetes, MLflow, Prometheus, Render, Azure ML, Google Cloud, AWS, CI/CD BI & analytics: Power BI, Tableau, Excel, Altair, Power Apps, Power Automate, A/B testing, Time series forecasting

More work

US Bancorp volatility: Holt-Winters and ARIMA failed Ljung-Box. ARIMA(1,0,1) + sGARCH(1,1) passed (p = 0.2369) and gives a 365-day price and volatility forecast from 11,835 daily prices. Stack: R, forecast, rugarch. StrataForge: Multiple regression for concrete strength (R² 81.35%, all five assumptions tested), turned into a worked £21,600 saving on a 2,000 m³ project. Stack: R, Hypothesis testing. Reder churn: Flattens nine nested JSON columns into a 12,483 × 55 table, adds RFM features, selects by mutual information and serves the model through FastAPI. Stack: Python, scikit-learn, FastAPI. Databricks big data: Spark SQL over 572,935 clinical trial records, and an ALS game recommender tuned by grid search and tracked in MLflow. Stack: Databricks, PySpark, Spark SQL, MLlib, MLflow. SQL Server design: A 9-table 3NF banking database with salted password hashes, procedures and triggers, plus NHS prescribing analysis with window functions and z-score outliers. Stack: SQL Server, T-SQL. IMF global economy: A Power BI star schema over 180+ countries and 44 IMF indicators from 2001 to 2020, with custom DAX. Stack: Power BI, DAX.

Every metric on this page comes from the linked repo, except Rebel Hive, which is private client work.


Website: brightuzosike.dev Email LinkedIn Portfolio

Pinned Loading

  1. reder-analytics reder-analytics Public

    A Predicting Customer Attrition project which aims to reduce customer churn for Reder company by leveraging predictive analytics and machine learning.

    Jupyter Notebook

  2. strataforge-prediction strataforge-prediction Public

    StrataForge’s predictive analytics project modelling concrete compressive strength using regression and hypothesis testing. Explores how mix composition (cement, water, slag, age, superplasticizer)…

    JavaScript

  3. data-science-projects data-science-projects Public

    Data science portfolio spanning finance, construction, telecommunications, and macroeconomics — featuring regression modelling, time series forecasting, ML classification, and interactive dashboards.

    Jupyter Notebook

  4. alloy-tower-data-platform alloy-tower-data-platform Public

    Live full-stack ML system for property valuation — FastAPI + Streamlit + MLflow + Docker, deployed on Render.

    Jupyter Notebook

  5. voyage-rails-analytics voyage-rails-analytics Public

    Dynamic-pricing ML service: gradient boosting + SHAP interpretability + MLflow tracking, containerised with Docker Compose.

    Jupyter Notebook

  6. sql-database-projects sql-database-projects Public

    TSQL