ADITYA FORECASTS
I build models that put a number on what happens next — who wins the tournament, which passage answers the question — and the cloud plumbing that keeps them answering after the demo is over.
Knockout-stage accuracy on the PMGC winner prediction model, over held-out tournament rounds.
Self-reported · method in write-up- 42% Modelling
- 26% Cloud & data
- 18% Writing & teaching
- 14% Experiments
Model card
- Trained on
- An engineering degree, then three years of building things that had to survive contact with real data.
- Optimises for
- Models that stay honest about their own uncertainty, and pipelines someone else can pick up without a phone call.
- Inputs
-
- Python
- PyTorch
- scikit-learn
- XGBoost
- FastAPI
- Azure
- Docker
- SQL
- Known limits
- Strongest on tabular and retrieval problems. Still growing the distributed-training and MLOps-at-scale side of the work.
- Status
- Open to machine learning and cloud engineering roles, and to collaborations on prediction problems.
Selected work 3 of 4
- 2025 PMGC Winner Prediction Ranking 16 teams by their chance of taking a PUBG Mobile Global Championship round, from historical placement and engagement data. XGBoost · scikit-learn · pandas 78% accuracy
- 2026 FIFA 2026 Bracket Simulator A browser app that plays the tournament forward thousands of times and reports how often each side survives each round. React · Vite · Monte Carlo Live demo →
-
2025 RAG Document Assistant Question answering over uploaded PDFs, with retrieval scored and cited so the answer can be traced back to a page. Azure AI Search · FastAPI Write-up in progress
Recent notes
Understanding generative AI
How VAEs, GANs and diffusion models each answer the same question differently: what does it mean to sample from a distribution you have never seen written down?
Read →What is a vector database?
Why similarity search replaced exact matching once we started storing meaning instead of strings, and where that leaves your existing database.
Read the notes →Get in touch
If you are working on something that has to predict, rank or retrieve — or you just want to argue about calibration — I would like to hear about it.