Software platform for an instrumented lab-animal IVC cage: FastAPI + React + Arduino/Raspberry Pi, closed-loop volumetric water dosing, and an honest account of why the reported 0.996 F1 does not m…

Grounded QA over the EU AI Act where the evaluation harness is the deliverable: retrieval, faithfulness and hallucination metrics over 45 hand-written questions

Vision-language hallucination measured against a hand-verified adversarial set. BLIP hallucinates 4.5% but misses 29% of present objects; CLIP verification cuts hallucination 33-40% at a real cost …

Time-series evaluation protocol, measured: the horizon inflates a forecasting score 10x more than the train/test split, and refitting daily for a month is worth 0.3%. Both premises this project was…

Anomaly detection for visual defect inspection: trained on normal images only, with localisation. PyTorch + MVTec AD.

Offline recommender metrics vs true online lift, graded against a known answer (Open Bandit Dataset). Ends in an evaluation gate that refuses to report a number when the diagnostics say it would be…

Fairness audit of my own fraud model against the EU AI Act - which turns out to expressly exempt fraud detection from high-risk classification.

A/B testing and causal inference scored against known answers: simulations where I set the effect, and a randomised benchmark the observational methods have to recover.

ARC-AGI-2 attempt: object-centric DSL with a verifier-backed program search. 3.9% on public training, 0% on public evaluation -- and a characterisation of why.

M4 forecasting where the prediction interval is the deliverable: nothing reaches nominal 95% coverage, measured residuals beat analytic bands at 29% narrower width, and seasonal naive wins outright.

Dark-theme browser calculator with operator precedence, parentheses and full keyboard support. No dependencies.

LoRA fine-tuning a small LLM for structured extraction, with the catastrophic-forgetting check most projects skip. Both numbers reported.

Custom Gymnasium arm environment + PPO, where the README leads with the two reward functions the agent exploited: deliberate self-termination, and farming the drift in potential-based shaping.

Benchmarks six data-drift detectors on harm-aligned precision/recall against measured model degradation, not on whether a distribution moved

Does an LLM narrating a GNN explanation describe the subgraph it was given, or the answer it was told? Perfect citations, wrong explanations.

A planted-confound benchmark for unsupervised anomaly detection -- built to prove heatmaps go to the wrong region, and refuting its own premise instead.

Node-level AUC, ring recovery and GNN explainer faithfulness are three different things: measured on synthetic fraud graphs against a random-edge null

Audit kit that measures how much reported GNN node-classification accuracy comes from transductive leakage and duplicate nodes, with a density control that shows most of it does not

Working IEEE-CIS Fraud Detection end to end, where the documented decision trail is the deliverable, not the leaderboard score.

Fraud model behind FastAPI, CI gating and drift monitoring -- where the monitoring is measured against a healthy null, not just demoed against an injected failure.