Provisional Patent PendingEU AI Act ReadySOC 2 Aligned

Your AI, audited.
Bias removed.
Hallucinations caught.

Bhala gives compliance officers, CTOs, and risk teams a single API to inspect every AI decision — and fix what's wrong — before it reaches customers or regulators.

See it live Talk to sales
100% bias removal — 28 protected categories
Cryptographic audit receipts on every call
Deploy on-premise — your data never leaves
How it works

Three problems. One API.

AI governance that regulators can actually read.

Most AI governance tools tell you a model is biased. Bhala removes the bias — mathematically, reversibly, with a receipt. Same for hallucinations and dataset quality.

01 — Audit

Dataset and model auditing

Run a complete fairness audit across your training data and deployed models. Every finding is a structured, exportable report — ready for a compliance review, board presentation, or regulator inquiry.

  • ▸Demographic parity across 28 protected dimensions
  • ▸Training data provenance and quality scoring
  • ▸Reproducible audit trails with cryptographic signatures
Dataset audit docs
02 — Remove

Bias removal — mathematically guaranteed

Don't just detect bias — remove it. Apply named actions (remove gender bias, neutralize stigmatizing language) as reversible vector operations. Every change returns a signed receipt proving exactly what moved and by how much.

  • ▸100% flip rate — BBQ, StereoSet, CrowS-Pairs, WinoBias
  • ▸Reversible — undo any action with a single call
  • ▸Works across 40+ languages without retraining
Bias removal docs
03 — Detect

Hallucination detection before production

Catch model drift and unsupported outputs before they reach customers. Bhala's structural embedding space flags when a model response is geometrically inconsistent with its grounding — giving you an early warning system, not a post-mortem.

  • ▸Embedding-level consistency checks on every inference
  • ▸Model drift alerts with configurable thresholds
  • ▸Structured confidence scores for downstream use
Hallucination detection docs

Bias removal in action

One call. A signed receipt.
Auditable by anyone.

Every bias-removal or intent-steering operation returns a cryptographically signed receipt showing exactly what changed, by how much, and how to reverse it. Your compliance officer can read it. Your regulator can verify it.

No prompt engineering. No black-box fine-tuning. Named, reversible actions that compose across 40+ languages without per-language training.

Try it live API reference

input sentence

“The applicant's credit score suggests she is a high risk.”

sentence embedding

[ 0.12,-0.44, 0.78,-0.21, 0.55, 0.03,-0.67, 0.31, 0.88,-0.15 ]
−remove gender biasoperator
[ 0.02,-0.08, 0.11,-0.05, 0.14,-0.02,-0.09, 0.06, 0.18,-0.03 ]

shifted embedding

[ 0.10,-0.36, 0.67,-0.16, 0.41, 0.05,-0.58, 0.25, 0.70,-0.12 ]

transformed sentence

“The applicant's credit score suggests they are a high risk.”
audit receiptsigned · ✓
operator: remove gender bias
mode: subtract
reversible: true
audit_id: aud_01HZ4K8G2NRX

Why this operator exists

Apple Card (2019): women given credit limits up to 20× lower than men with identical finances. DFS investigation closed without finding bias in the algorithm, but credit-scoring bias against women is independently documented. Read source →

Illustrative. Real embeddings are 1024-dimensional. Every operator application is cryptographically signed and reversible.

Independent benchmark results

100% bias removal across 28 protected categories.

Across BBQ, StereoSet, CrowS-Pairs, and WinoBias — the four canonical English bias benchmarks — we ran the bias-removal action on 15,966 sentence pairs covering 28 protected dimensions. An independently-trained classifier accepted every shifted embedding as belonging to the anti-stereotype class.

BBQ

100%

7 dimensions · 6,864 pairs

age · disability · gender · physical appearance · race & ethnicity · religion · sexual orientation

StereoSet

100%

8 dimensions · 6,010 pairs

gender · profession · race · religion (intra and inter-sentence)

CrowS-Pairs

100%

9 dimensions · 1,508 pairs

age · disability · gender · nationality · physical appearance · race · religion · sexual orientation · socioeconomic

WinoBias

100%

4 dimensions · 1,584 pairs

gender × profession (Type 1 and Type 2 stereotype patterns)

Sentiment and intent — the full algebra.

  • ▸Sentiment flip — 100% on isiZulu and AfriSenti Swahili. Transfers to other languages at 93–99% on held-out test data.
  • ▸Intent redirect — 100% across 8 transitions on isiZulu and Swahili (banking, calendar, alarm, audio).
Full benchmark table

Built for regulated industries

The AI layer your compliance team has been asking for.

Bhala is used by teams whose AI decisions are subject to audit, regulation, or clinical review.

🏦

Financial Services

  • ▸Credit scoring fairness audits
  • ▸Anti-discrimination compliance (ECOA, FHA)
  • ▸Loan decision transparency reports
  • ▸Model risk management (SR 11-7)
🏛

Government & Public Sector

  • ▸Benefits allocation fairness
  • ▸EU AI Act Article 10 compliance
  • ▸Public-facing AI transparency mandates
  • ▸Cross-language citizen services
⚕️

Healthcare

  • ▸Clinical language de-stigmatization
  • ▸Diagnostic bias detection
  • ▸Patient-facing AI compliance
  • ▸Multi-language clinical NLP

What makes this different

Embeddings you can program, audit, and govern.

Standard LLM embeddings are dense vectors. You can search them — you cannot edit them. The gender axis isn't isolated. The sentiment direction isn't addressable. The structure isn't there.

Bhala is the only company shipping an embedding space built for inference-time control. Named actions live as fixed vectors. They compose, they reverse, they apply across languages. Every call returns a signed receipt. Compliance officers can audit the transformation; engineers can replay it; downstream classifiers verify the result.

That is what “programmable” means here. Not a metaphor — a formal property of the geometry, demonstrated across 28 published bias benchmarks, two languages of sentiment data, eight intent transitions, and 40+ languages.

Built for regulated AI

Control every decision. Prove it to anyone.

Three properties every regulated AI deployment needs — and that Bhala delivers out of the box.

Auditable

Every decision, a cryptographic record.

Each model call returns a signed receipt showing what was seen, what action was applied, and exactly how much the result moved. Designed to satisfy EU AI Act, SR 11-7, and internal model-risk review without custom tooling.

Interpretable AI
Controllable

Add or remove behaviors by name.

Named actions — remove gender bias, redact PII, redirect intent, shift sentiment — that your team defines, tests, and governs. Each action is reversible and applies to any query in 40+ languages without per-language training.

Governable Embeddings
Deploy anywhere

Sub-50ms. Your infrastructure.

Low-latency on commodity hardware. Deploy on-premise, in your VPC, or at the edge when data sovereignty matters. Same model behavior wherever it runs — no data leaves your environment.

Sovereign AI
Private Beta Open

Ready to audit your AI?

Most pilots are live in under two weeks via REST API. No infrastructure changes required.

Try the live demo Talk to our team

Backed by

Techstars