Michal Kaszubski

NeuralTake neuraltake.com (external link)

Founder and AI engineer · London

• Secured and delivered more than £5,000 in signed B2B AI software and consultancy contracts (signed value, not revenue).
• Owned stakeholder discovery, scoping, requirements, architecture, delivery, onboarding, support and roadmap.
• Designed and deployed a production AI SaaS (AI Zuzi) in Django, React and PostgreSQL: AI-assisted generation, multilingual workflows, Stripe payments and seven third-party integrations (Facebook, Instagram, LinkedIn, X, TikTok, YouTube, Pinterest) plus WordPress blog publishing; used by approximately 50 business users.
• Took a client platform from prototype to production through a security audit, encryption at rest, MFA, GDPR-aligned deletion and a zero-data-loss live SQLite-to-PostgreSQL migration.

NeuralKite neuralkite.com (external link)

Product design and engineering · London

• Acquisition-target intelligence for fragmented UK private markets: ownership discovery, screening, thesis-fit scoring, shortlisting and export, producing decision-ready dossiers.
• Next.js App Router and TypeScript, Prisma and PostgreSQL, NextAuth, Stripe billing, LLM agents on the Anthropic SDK, Playwright and Cheerio scrapers, SSE streaming for live discovery.
• Self-hosted PostgreSQL warehouse including a 2.8-million-company Crunchbase dataset, loaded and analysed.
• Deployed and publicly accessible; customer and usage figures not published.

ChefBot

Designer of retrieval, prompt logic and tool flows

• GPT-based cooking assistant that helps people make meals from sparse fridge or pantry inputs; built on GPT (orchestration, not model training).
• Designed the retrieval, prompt logic and multi-step tool flows and iterated the experience using usage patterns and user feedback.

Automated detection of geometric distortion in prostate diffusion-weighted MRI

MSc thesis, UCL github.com/Cashubski/miqa-prostate-dwi (external link)

• Retrospective dataset of 1,027 acquisitions from 627 patients (single centre, anonymised); strict patient-level five-fold cross-validation.
• Three-seed DenseNet-121 ensemble: mean AUROC 0.819, SD 0.049 across the five folds. Heterogeneous nine-model ensemble: uncertainty-error AUROC 0.685; 83.6% retained-case accuracy at 70% coverage in a retrospective triage simulation.
• Distilled a single model that reduced forward passes from nine to one. PyTorch, classical baselines, bootstrap confidence intervals, paired Wilcoxon tests, Grad-CAM, calibration analysis, synthetic-distortion validation; public code under MIT.

RSNA Knee Abnormality Detection (Kaggle research code competition)

Ongoing entry www.kaggle.com/competitions/rsna-knee-abnormality-detection (external link)

• Weak supervision: 58 of 4,407 training studies carry the 12 gold labels, every study has a free-text radiology report; pipeline report to labels to image model to image-only inference.
• timm slice encoder, BiGRU, attention pooling and series aggregation over 2.5D slice stacks; DistributedDataParallel on two Quadro GV100 GPUs, patient-level GroupKFold, BCE, AMP, cosine schedule.
• About 500 GB of raw DICOM preprocessed into windowed uint8 shards at 384 px on Kaggle and synced to a GPU cluster. Ongoing as of 15 September 2026; no result yet.

MSc, Artificial Intelligence and Medical Imaging

University College London

BSc, Computer Science

City, University of London

PyTorch, timm, DistributedDataParallel, deep ensembles, MC dropout, selective prediction, calibration, SimCLR/BYOL, Grad-CAM, GroupKFold, DICOM and NIfTI pipelines

Django, React, Next.js, TypeScript, Prisma, PostgreSQL, NextAuth, Stripe, Anthropic SDK, Playwright, Cheerio, SSE, pytest

Discovery, scoping, requirements, architecture, onboarding, support, roadmap; security audits, encryption at rest, MFA, GDPR-aligned deletion

• Cash Nova: a fictional producer created to explore generative music; 52 tracks across six albums on ElevenMusic, 97 followers (15 September 2026); tracks received platform discovery and playlist placement.
• CleverFacts and Ogarnik: one short-form educational format adapted for English and Polish audiences; nearly 720,000 combined views with approximately 1,000 subscribers on each channel (15 September 2026; views, not unique viewers).