# Vet Microscopy AI — vet-ai.ucognet.pro
> AI-assisted parasite egg microscopy with calibrated confidence, anomaly detection and human-review routing. A Brainstream product, built on the U-CogNet research platform. 11-species classifier (Chula-ParasiteEgg-11), 97.36 % held-out accuracy with 95 % CI [96.24, 98.16], full audited safety stack, founder pilot from €2,500.

## What this site is

This is the COMMERCIAL surface of Vet Microscopy AI by Brainstream. The product is built on the U-CogNet research platform, where peer-reviewed methodology, full per-experiment audits, scientific references with DOIs, and the original honest-caveats discussion are openly published. See https://ucognet.pro/researches/vet for the research view of the same model.

## Headline result (audited on the public Chula-ParasiteEgg-11 corpus)

- Held-out accuracy: 97.36 percent on 1,100 test images (1,071 of 1,100 correct).
- 95 percent confidence intervals: Wilson [96.24, 98.16]; Bootstrap (5,000 resamples) [96.36, 98.27].
- macro-F1 = weighted-F1 = 0.974. Per-class recall in [94, 100] percent across all 11 species; three species at 100 percent recall (Fasciolopsis buski, Hookworm egg, Hymenolepis diminuta); lowest is Ascaris lumbricoides at 94 percent.
- Coverage / accuracy abstention: at threshold tau = 0.90 the system commits on ~71 percent of cases at ~99.5 percent accuracy and routes the remaining ~29 percent to a human reviewer.
- Calibration: Expected Calibration Error 0.074 -> 0.020 (-72 percent) via per-class temperature scaling.
- Unsupervised anomaly channel: free-energy autoencoder trained ONLY on Ascaris flags the other 10 species at AUROC 0.82-0.99 (mean 0.917).
- Leakage audit: ResNet18 embedding cosine nearest-neighbour test->train flags 0 / 1,100 images at cos >= 0.98 (max observed 0.952). Perceptual-hash flags 311 / 1,100 at Hamming < 8, which we attribute to shared microscope and staining (not memorisation); both views are published openly.

## What we sell

Pilot programme — €2,500 to €7,500, founder pricing, first 3 design partners.

Deliverables:
- Fine-tuned classifier on the customer's dataset (or our public-data baseline if they do not yet have labelled data).
- Full v0.2 defensive audit (confusion matrix, per-class precision/recall/F1, accuracy with 95 % CIs, leakage check).
- Per-class temperature calibration plus reliability diagram.
- Coverage-accuracy abstention curve at the customer's chosen operating point.
- Free-energy unsupervised anomaly channel (second safety view).
- Grad-CAM attention maps for every prediction.
- Evidence Manifest per run, bit-exactly replayable from disk alone.
- Integration design plus one month of operational support.

Target audiences:
- Veterinary diagnostic laboratories
- University parasitology / One-Health programmes
- Microscopy education
- Quality-control screening for animal diagnostics

Not in scope for v0.1:
- Human-clinical deployment (would require full DURC / IBC review, FDA pathway).
- Synthetic-biology screening (that is the separate Biomolecular Safety project — see ucognet.pro/researches).

## How it works

1. Classify: a fine-tuned ResNet18 backbone classifies the microscope image across 11 species. Per-species precision and recall are reported individually.
2. Calibrate: per-class temperature scaling (Guo et al. 2017) reduces ECE from 0.074 to 0.020.
3. Escalate: below a configurable confidence threshold, the system abstains and routes the image to a human reviewer. A second free-energy autoencoder channel flags OOD inputs independent of the supervised classifier.

The classifier sits inside the full 45-module U-CogNet integrated cognitive system. The safety primitives (per-class temperature calibration + free-energy autoencoder + L2 class-aware drift guard + coverage-accuracy abstention) constitute the Tier-3 safety stack. Same architecture proven cross-domain on patient-aware MIT-BIH cardio arrhythmia detection.

## Honest scope and caveats

- The 97.36 percent accuracy is on the Chula-ParasiteEgg-11 distribution. Cross-laboratory generalisation requires the external-style robustness battery (which is published at the research site and pending a clean end-to-end run) and ideally an independently-collected held-out lab.
- The cognitive moat is in the safety stack, not the backbone. Any competitor can train a ResNet18 to similar accuracy on this corpus. The differentiator is calibrated abstention, free-energy escalation, the L2 drift guard, and the bit-exact Evidence Manifest — all of which fail loudly and are evaluator-defensible.
- We deliberately disclose the perceptual-hash leakage flag (311 / 1,100) because evaluators will ask; the discriminative embedding check (0 / 1,100 at cos >= 0.98) is the answer.

## Contact

- Pilot enquiries: vet-ai@brainstream.pro
- Founder direct:  samuel@brainstream.pro
- Organisation:    Brainstream (https://brainstream.pro)
- Research site:   https://ucognet.pro
- Vet research tab: https://ucognet.pro/researches

We respond within 48 hours.