Skip to content
balisa_

Open to internships · Fajikunda, The Gambia

BSc Statistics, KNUST · Mastercard Foundation Scholar

Abdoulie Balisa

Statistics student. I build the model and the thing that serves it.

I mostly work on retrieval and forecasting systems, and on the checks around them. A lot of what I have learned came from getting those checks wrong first: answers that cited sections which did not exist, and validation splits that flattered the model.

  • 9Shipped
  • 2In development
  • 2024Shipping since

Selected work

Three projects, in full

Independent research, a synthetic-data engine, and a retrieval system over Gambian law. Each row links to the live product, the code where it is public, and the case study.

This projection reaches 4.66 million by 2074, about 0.7 million under the UN medium variant. It is based on the 2024 census; the UN's was finalised before it.

This projection reaches 4.66 million by 2074, about 0.7 million under the UN medium variant. It is based on the 2024 census; the UN's was finalised before it.

01Live

The Gambia 2074

An independent population forecast for The Gambia, out to 2074

An independent population projection for a country whose published figures were set before the 2024 digital census. 7 million below the UN. The engine reproduces the UN's own projection to within 1 percent when fed their inputs. Three forms of Lee-Carter mortality, run over 1,000 futures.

~4.66M by 2074 (4.35 to 4.98M), within 1% of the UN

  • Python
  • PyMC
  • MCMC
  • NumPy
  • Pandas
  • Matplotlib
CodeArticleCase study
The four steps from describing your columns to downloading the CSV, and the four checks every batch has to pass.

The four steps from describing your columns to downloading the CSV, and the four checks every batch has to pass.

02Live

NOVA

A synthetic-data engine for finance, from domain rules or from real data

Banks in West Africa hold customer data they cannot share, and for rural borrowers the data was never collected. NOVA generates synthetic data two ways: from a description of the columns and their rules, or from a real dataset via a Conditional Tabular GAN written from scratch in PyTorch. 10.

TSTR 0.92 · 4 checks pass · 7 domains, no source data needed

  • Python
  • PyTorch
  • CTGAN
  • FastAPI
  • Next.js 16
  • scikit-learn
LiveCodeCase study
Ask in plain English and the answer names the Act and section it came from.

Ask in plain English and the answer names the Act and section it came from.

03Live

Gambia Legal Aid

RAG chatbot for Gambian law

Lawyers in The Gambia are expensive, the statutes are scattered across PDFs, and a hallucinated answer in a legal question does real harm. I built a RAG system over 13 Acts of Parliament where answers name the section they came from, or say the corpus does not cover the question.

Cites the section, or refuses. 13 Acts, validated before it answers

  • Python
  • RAG
  • Vector search
  • FastAPI
  • Next.js
LiveCode privateCase study
See all 11 projectsRead the case studies