Tangeni Idhenga

Tabular

Turns a bank statement PDF into a clean CSV, Excel or OFX file.

It checks every row against the balance printed on the statement and tells you which ones don't match.

Text PDFs go through a deterministic parser and only scans go to Gemini's vision model. Every row is reconciled against the statement's own running balance in integer pence, so a figure that doesn't add up gets flagged instead of slipping through. An accuracy harness runs in CI, a privacy test fails the build if bank data touches disk, and I patched React2Shell (CVE-2025-55182) on the live deployment.

Next.js 15, React 19, FastAPI, Gemini, Turborepo, pytest

ClearPass

A tool that attacks AI assistants and tells you which attacks worked.

It tries prompt injection, data leaks and tool abuse, then gives every attack a pass, fail or unsure and explains how it got there. When I ran it on Google's gemini-3.1-flash-lite, the model read out its own system prompt and repeated a fact poisoned into a document.

Attacks are YAML files mapped to OWASP and the EU AI Act. Rules look for direct evidence first, and if they find none, a second LLM reads the exchange and decides. The target's reply is hostile input, so each judge call uses a random nonce and a garbled answer counts as unsure. Live testing showed a bad API key could produce a clean 100/100, so an untested system now gets unsure and no score. I also found and fixed a stored XSS in the report, and the API has an SSRF guard, encrypted keys and an authorisation gate before any probe fires.

Python, Pydantic, FastAPI, SQLAlchemy, Next.js, TypeScript, Tailwind, Gemini, pytest

Oil supply chain resilience under geopolitical shock

For my MSc dissertation I built a simulation of the oil supply chain, calibrated it to real 2026 prices and tested it against the Strait of Hormuz closing.

I wanted to know if a field's position in the supply chain matters more than its geology. The field with the best rock and the cheapest barrels came last because all of its oil goes through the Gulf.

I built five agent types in Mesa 3 and calibrated the price mechanism against real EIA Brent and WTI data. A LASSO analysis found route exposure was the biggest driver of vulnerability, ahead of any geological property. The ranking stayed the same across five seeds, three onset dates and over a thousand changes to the scoring weights.

Python, Mesa 3, Monte Carlo, LASSO, Next.js

Gradaro

A property search for first-time buyers that starts from what you can actually afford.

It works out your buying power, stamp duty and monthly cost before it shows you any listings, then flags things like short leases and flood risk.

It uses affordability and valuation engines in TypeScript, Leaflet maps and a Python valuation pipeline. I built it under ILP Solutions and I'm still working on it. The listings in the demo are sample data.

React 18, TypeScript, Tailwind, Leaflet, Python

Sound × AI

A generative music studio that makes original techno and runs fully in the browser.

Pick a style and it writes a new track. You can also drop in your own song and it picks up the tempo, key and energy.

The melody model is a GRU I trained from scratch in NumPy, with the forward pass, backprop through time and Adam written by hand. I then rewrote the forward pass in TypeScript so it runs on the client with no backend and no ML runtime. It exports to WAV, stems and MIDI.

React 19, TypeScript, NumPy, Tone.js, Web Audio

Exposing security vulnerabilities in open-source AI

For my BSc dissertation I tried backdoor attacks on eight open-source models from Hugging Face, across audio, image and text.

None of the attacks worked reliably. One looked like it had, but it was a bug in my test harness.

Single-shot poisoning didn't reliably work on any of the eight, and the text generators looked the easiest to attack. I later rebuilt the code as a Python package with tests and a CLI, so everything can be rerun.

Python, PyTorch, Transformers, pytest

Co-founder and software engineer, ILP Solutions

May 2024 to now, London

• I lead the technical side and delivery for our clients, so far that includes the Namibian High Commission and the African Union's London diaspora project.
• I'm building Gradaro, a property search for first-time buyers, in React and TypeScript with a Python valuation pipeline.
• I rebuilt the Namibian High Commission website in React and Tailwind and worked on accessibility and load times.

Software engineer, Afrivest, fintech startup

Jun to Aug 2024, London

• I was the only developer on the MVP of an investment platform and worked directly with the CEO.
• I wrote a custom plugin that pulls live market data from third-party APIs into the alpha, and retention went up by about 20%.

Systems engineer, placement, Namibian High Commission

Feb to Aug 2024, London

• I built an access-controlled database to digitise the Commission's paper records, which cut processing time by around 20%.
• I rewrote the public website's content and UX, and support queries dropped by around 70%.

Software engineer, placement, Taveo, insurtech startup

Jun to Dec 2023, London

• I built a Microsoft Teams bot that transcribes and summarises meetings using Azure, AWS Lambda and the OpenAI API. It cut manual note-taking by 60% across the company and I demoed it live to broker partners.
• I worked out the data integrations with over ten UK insurers. I also used Matomo analytics to decide what to change on the broker web app, and functionality scores went up by 15%.

Project manager, Bonny Vainö Idhenga Foundation

Dec 2020 to Apr 2024, Windhoek, Namibia

• I led 14 volunteers supporting Himba children in northern Namibia.
• I ran donation drives for local orphanages during the Covid lockdowns and rebuilt the foundation's website.