Data & ML

Fyregig

Sports Prediction & Analytics Platform

Custom ML model trained on 80,000+ combat-sports records delivering 80%+ prediction accuracy

01

The problem

The client wanted high-accuracy outcome prediction for combat sports: a domain where public datasets are fragmented, inconsistent and incomplete. There was no usable training corpus to buy, and the project carried tight budget constraints, so the work had to deliver research-grade accuracy without research-grade cost.

02

How we built it

6 decisions
  1. Built an in-house data pipeline that scraped, normalised, de-duplicated and cleaned over 80,000 historical fight records into a consistent training corpus.

  2. Engineered domain-specific features: fighter attributes, style matchups, historical performance, streaks, physical differentials and contextual variables.

  3. Trained, tuned and validated a custom machine-learning model, achieving 80%+ accuracy on upcoming matches.

  4. Built a responsive interface presenting predictions, statistics and insights in a form a non-analyst can act on.

  5. Added advanced filters, historical comparisons and performance indicators that went beyond the original brief to improve engagement.

  6. Delivered the full scope within tight budget constraints by balancing model sophistication against compute cost.

03

System architecture

6 layers

Every layer designed, built and operated for this system to run in production.

L01

Machine Learning

Python: scikit-learn / gradient-boosting model family with tuned ensembles

L02

Data Engineering

Python scraping and ETL pipeline with cleaning, validation and versioned datasets

L03

Data Analysis

pandas, NumPy for feature engineering and exploratory analysis

L04

Model Serving

Python inference API exposed to the web application

L05

Frontend

React with a responsive analytics-oriented interface

L06

Hosting

Cloud-hosted model service with a separately deployed frontend

04

What was delivered

7 modules
  1. 01

    Data Acquisition Pipeline

    Automated scraping, normalisation, de-duplication and validation of historical records.

  2. 02

    Feature Engineering

    Domain-specific feature construction from fighter, matchup and contextual variables.

  3. 03

    Model Training & Validation

    Model selection, hyperparameter tuning, cross-validation and accuracy benchmarking.

  4. 04

    Prediction Service

    Inference API serving upcoming-match predictions with confidence indicators.

  5. 05

    Statistics & Insights UI

    Match predictions, fighter statistics and comparative performance views.

  6. 06

    Advanced Filters

    Multi-dimensional filtering across fighters, weight classes, events and time periods.

  7. 07

    Historical Comparisons

    Head-to-head and cohort comparison tooling.

05

Outcomes

4 results
  • 80%+ prediction accuracy achieved on upcoming matches: the client’s primary success criterion.

  • A proprietary, cleaned 80,000-record dataset created as a durable asset for the client.

  • Delivered inside tight budget constraints without compromising model quality.

  • Client specifically noted the model accuracy and the additional engagement features delivered beyond scope.

06

Value delivered

  • For the client: a differentiated data product built on an asset competitors cannot simply buy.

  • Demonstrates DevionX capability across the full ML lifecycle: acquisition, engineering, training, serving and UI.

Client and product names are used with reference to publicly available product information. Engagement details covered by non-disclosure agreements are described at capability level only.