Every claim, verifiable.

Detect bias by alignment, not absorption

In practice:The live /demo/audit endpoint computes the EEOC 4/5ths-rule disparate-impact ratio across six regulated domains (criminal sentencing, employment, online moderation, venture capital, police interaction, higher education) for any text you paste. Sub-second per audit, signed receipt. Predictions come from labeled-example projection, not from baked-in associations.

1. Algebraic de-racing. We encode your text once, then compute a neutralized control by subtracting the learned race-direction component (compose, not string substitution). No hardcoded name list — works on any text.
2. Constructor synthesizes per-domain race direction. A single learned linear constructor takes (T_race, T_domain_context) → T_race_in_domain. Held-out cos to ground-truth = 0.91 across 4 unseen domains — generalizes to any new domain without re-training. Same primitive that won on logic AND with held-out cos 0.999.
3. Disparate-Impact Ratio + 4/5ths rule. EEOC says DIR < 0.80 fails Title VII / ECOA. We report DIR per domain plus statistical-parity-difference (SPD) and a 0–10 risk score. Verdict cards say PASS / MARGINAL / FAIL · 4/5ths rule.
4. Intersectional via ZFC set intersection. Race × gender per domain via set-intersection on the SCI manifold. The audit cell EEOC explicitly requires.

Methodology:T_race / T_gender / T_religion operators trained on contrastive pairs from Bertrand-Mullainathan first names + US Census 2010 surnames (180K names total) embedded in 18 deployment-domain frames (10 frames per domain, grounded in audit literature). Race-in-domain constructor: 256-d linear over [T_race ⊕ T_dom_ctx] → 128-d L2-normalized; trained on 14 domains, evaluated on 4 held-out. Receipt = SHA-256(text + model_version + scores + timestamp).

Bias Removal

Detects and removes bias in AI text on demand. You ask the AI to remove a specific bias — gender, race, religion, age, disability, or any of 23 other categories — and it does, with an independent classifier confirming the change worked. No retraining, no per-category tuning.

Technical detail:100% correction rate across 4 industry-standard fairness benchmarks (BBQ, StereoSet, CrowS-Pairs, WinoBias) covering 28 protected dimensions and 15,966 sentence pairs. We apply a “remove bias” control at inference time; an independently-trained classifier verifies every shift. Reproducible in under 90 seconds on a laptop GPU.

Hate Speech Detection

Detects hate speech across 12 protected groups. Trained to flag hateful content on social media, news comments, and adversarial inputs — and crucially, to NOT flag people discussing hate (counter-speech), in-group reclamation, or news reporting on slurs.

Technical detail:AUROC 0.90 on HateCheck (adversarial templates), 0.77 on TweetEval-hate (Twitter, OOD). Trained jointly on 11 corpora (~134K labeled examples). 15M-param backbone + 528K-param adapter — 7× smaller than HateBERT (110M) or HateXplain BERT (110M). Live in production on bhala-api since 2026-05-02.

Sentiment Analysis

Two capabilities for sentiment. Sentiment steering — push AI output toward positive or negative tone using a single fixed vector, in any language, with an audit trail per shift. Sentiment reading — classify how positive or negative a piece of text is, even in languages with little training data.

Technical detail:Steering is inference-time control: 100% sentiment-flip success on every language tested in-family, 77% cross-family transfer to English (Bantu→Indo-European, zero-shot). We apply the same control to any language without per-language tuning. Reading is AfriSenti classification: 89–91% of fine-tuned SOTA on 5 African languages with a frozen 15M-param backbone (vs. 270M+ baselines).

Intent Classification

Can the AI tell what the user wants? Book a flight, set an alarm, request a refund, transfer money — these are intents. We tested across 51 languages and 60 different intents, and matched or beat models 100× our size.

Technical detail:We test the strict way: freeze the model, then train the simplest possible classifier — a single linear layer — on its output. A high score means the model already did the hard work of organizing the concepts; the classifier itself adds almost nothing.

On MASSIVE (FitzGerald et al. 2022 — 51 languages × 60 intents), the frozen 15M-parameter encoder scores 73.2% Swahili (above GPT-4o's 70.6%), 72.5% Korean, 69.7% Hindi, 66.5% Amharic — 38–43× over chance, with none of these languages anywhere in the pretraining data. We also report Injongo (8-language Bantu intent benchmark, with published SOTA comparison).

Why a linear probe, not something stronger? We also tried a more powerful classifier — a 2-layer neural network — and it did worse on all three cross-family transfers (Korean / Hindi / Amharic). That is the result: when the simplest classifier beats the stronger one, the concepts are already separated by straight lines in the model's geometry — genuinely organized, not just buried in there waiting for a powerful classifier to untangle.

Cross-Lingual Transfer (NER and beyond)

The intent results above (12 languages, 7 families) are the headline cross-lingual finding. Below: structural transfer also extends to other tasks — Named Entity Recognition without per-language fine-tuning.

Technical detail:Zero target-language training. Next-word prediction transfer: Nguni sister languages 87.5%, broader Bantu 71.1%, non-Bantu African 75.8%. Cross-family zero-shot: Yoruba 80.5%, Igbo 77.7%, Hausa 69.3% (Afro-Asiatic), Swahili 69.1%. Morphology-aware script handling unlocks 9 writing systems (Arabic, Hangul, Ge'ez, Devanagari, etc.) with no encoder retraining. Per-language adaptation builds in under 2 seconds.