PLAYER 1 · READY
Chukwudi Eke
Physician · Independent ML researcher
Open to data science and ML roles · remote-friendly
I treat models like patients: pre-register the trial, then report what it found.
Evaluation and calibration for medical-imaging models, world models, and football analytics. I care less about whether a model looks good than whether its confidence can be trusted when the data moves.
QUEST LOG
Showing 13 of 13 projects
LoRA lifts in-domain AUROC 0.588 → 0.775 but only 0.650 → 0.690 on a new robot. An auto-accept threshold validated at 1.8% error runs at 26.5% on deployment.
| Setting | Auto-accept rate | Realised error | Versus target |
|---|---|---|---|
| Same robots (RLBench + WidowX) | 14.5% | 1.8% | within target |
| New robot (UR5, never seen) | 7.9% | 26.5% | 5.3 times over |
GitHub
HFScale stops models over-calling tropical disease, but under-calling of treatable cosmopolitan disease in a low-income framing rises with scale. Three frontier models under-call at 0.82–1.00. Not scale, not medical fine-tuning, not the frontier: there is no safe tier.
Mid-layer geometry localises lesions without labels across CT and ultrasound, MedDINOv3 and DINOv2. Membership pruning beats saliency on small-lesion miss-rate, with a conformal retention certificate and 1.6× fewer FLOPs.
Scans 11M+ biomedical papers and live regulatory sources for safety signals, interactions, patient-specific dosing, pharmacogenomics, and Africa formulary status (NAFDAC · SAHPRA · WHO).
A biologically inspired CNN with learnable Gabor filters and competitive inhibition, trained with a differentiable calibration term (CE + λ·ECE) and confidence-gated escalation.
Fuses per-modality signals, escalates to a clinician on contraindications, hash-signs every output for audit, and degrades gracefully when a modality is missing. Runs in the browser, no install.
Finds exploitable space by running persistent homology over a velocity-aware pitch-control field. Validated as a shot precursor at AUC 0.886.
Every claim carries a machine-readable scientific tier so the interface can't over-claim, and signing impact is simulated live through a validated tactical twin.
An XGBoost ensemble, a Dixon-Coles goals model and a Bi-LSTM sequence model behind a FastAPI service, covering outcome, goals, both-teams-to-score, correct-score and handicap markets.
WaveletAdam solves noisy XOR 10/10 against Adam's 7/10 at σ = 0.05, with 6× lower final loss. Pure NumPy, 53+ tests.
The axes are measured near-independent, so matches are reported per axis, never as one style score. Descriptive, not a ranking.
READING ROOM
Clinical imaging is where calibration stops being academic.
I trained as a physician before I trained models. On a CT read, a confident wrong answer has a patient attached, so my imaging work is about what a model actually encodes and whether its confidence survives a new scanner, site or population.
- 3D CT foundation models · DIRAC, parameter-matched against CNN and ViT on KiTS23, MSD and AbdomenAtlas
- A locked CT pool · 6,212 abdomen and chest cases, QC'd, deduplicated, split with zero leakage
- Label-free SSL · covtoken, coverage-constrained token pruning for medical images
- Lung CT · probing frozen MedDINOv3 features for lesion identity on RIDER
RULES OF PLAY
- 01
Write the success criteria before the data can argue back.
- 02
A null result reported beats a positive one massaged.
- 03
Match parameters, or it's a capacity result wearing a structure costume.
- 04
Probe on CPU before spending GPU.
CONTINUE?
Building something that needs its confidence to hold up? Player 2 welcome.
Open to data science and ML roles · remote-friendly