Michal Kaszubski

AI/ML engineer and founder

I investigate how models fail, build production AI systems and create products that people actually use.

To get in touch, message me on LinkedIn (opens in a new tab). CV and references are available on request.

The evidence, in context

Three figures from the work below. Each is labelled by the kind of evidence it is, and dated where it changes over time.

1,027
MRI acquisitions evaluated
Retrospective prostate DWI dataset, 627 patients, patient-level cross-validation, UCL MSc research.
Research evidenceThesis, October 2025 to August 2026
About 50
business users
Of AI Zuzi, a production AI platform I engineer for a client through NeuralTake.
Commercial validationStated September 2026
25,000+
ChefBot conversations
A cooking assistant I designed, rated 4.7 from 100+ ratings as a GPT; it continues as a ChatGPT plugin. Conversations, not people.
User adoptionGPT store listing, screenshots from September 2026

Selected works

Four works, each card saying what it is, what I did and what happened as a result. The index lists all eight, including the two experiments.

See the full index of works

Thesis bar chart of uncertainty-error AUROC, with error bars, for deep ensembles, MC dropout and test-time augmentation. ResNet-18 scores 0.64 to 0.69 and ResNet-50 0.60 to 0.65. The DenseNet-121 deep ensemble reaches only 0.52, just above the dashed chance line at 0.50.

Thesis figure: how well each model's uncertainty ranks its own errors. Chance is 0.50.

When the most accurate model did not know when it was wrong

Research, 2025 to 2026

MSc research at UCL on automated detection of geometric distortion in prostate diffusion-weighted MRI, and on whether a model can tell when its own judgement is unreliable.

Role: MSc researcher, UCL (thesis author)
Built with: PyTorch, DenseNet-121, Deep ensembles, Selective prediction
Evidence: Research evidence
Result: Three-seed DenseNet-121 ensemble: mean AUROC 0.819, SD 0.049 across five patient-level folds, up to 0.17 AUROC over classical radiomic baselines. Retrospective triage simulation: 83.6% retained-case accuracy at 70% coverage (thesis, 2026).

Inspect the work

The AI Zuzi Socials Manager headline on a dark blue ground, with the line Create. Schedule. Automate. and a description of an all-in-one platform for creating, scheduling and distributing posts across channels.

Screenshot, socials.aizuzi.com, September 2026

AI Zuzi: taking a client’s social-media platform from prototype to production

Products, 2026 to present

An AI social-media management and content platform that I engineer for a client company through NeuralTake. Django, React and PostgreSQL, with seven third-party integrations, generation, payments and multi-factor sign-in. About 50 business users.

Role: Lead engineer through NeuralTake: back end, front end, integrations, security, database migration and deployment
Built with: Django, Django REST Framework, React, PostgreSQL
Evidence: Production deploymentUser adoption
Result: Used by about 50 business users, as stated in September 2026; a zero-data-loss live migration from SQLite to PostgreSQL; a security audit, encryption at rest, MFA and GDPR-aligned deletion before strangers were given accounts.

Inspect the work

NeuralKite

Products

A live acquisition-intelligence product for fragmented UK private markets. At its centre is an email discovery and verification engine that grades every address and keeps its reasons. Relevant to AI engineering, private-equity and finance roles.

Role: Product engineer, end to end: verification engine, data layer, scrapers, Claude agents, application and sending
Built with: Next.js App Router, TypeScript, Anthropic SDK, PostgreSQL
Evidence: Production deployment
Result: Deployed and publicly accessible at neuralkite.com, verified 15 September 2026. Its verification engine grades every address into a risk tier with a written rationale. Customer and usage figures are not published.

Inspect the work

ChefBot

Products, 2024 to present

A cooking assistant that turns whatever is in the fridge into a meal. It began as a GPT (25,000+ conversations, rated 4.7 from 100+ ratings) and continues as a ChatGPT plugin with 18 skills.

Role: Designed the retrieval, prompt logic and multi-step tool flows; orchestration on GPT, not model training
Built with: ChatGPT plugin (skills), Prompt logic, Multi-step tool flows, Image input
Evidence: User adoptionProduction deployment
Result: 25,000+ conversations and a 4.7 rating from 100+ ratings as a GPT (store listing, screenshot September 2026; conversations, not people); continues as a plugin since October 2026.

Inspect the work

Also:NeuralTake, my consultancy;RSNA knee MRI challenge, closed October 2026;Open source, twenty-two merged pull requests.

Now

Contributing fixes to open-source quant libraries, twenty-two merged so far, and engineering AI Zuzi for a client through NeuralTake, my consultancy.