01
Intent extraction
Detects what the model is trying to do, the context type, and the signals that matter for the decision.
NDPR-native accountability for AI teams
Make every AI decision inspectable.
VERDANT wraps any AI call and returns a structured reasoning chain: intent, fairness baseline, bias flags, a plain-language explanation, and a trust score. Built for Nigerian and African builders shipping AI into high-stakes contexts.
5 stages
every AI call becomes inspectable
0 freestyle
structured JSON at every step
40 threshold
webhooks fire below this score
NDPR-native
audit logs built for local compliance
audit_id: vd_a7f3c91
Candidate Shortlist · Hiring
84
/ 100 trust
01
IntentHiring — candidate review
02
BaselineNG hiring baseline v2.1
03
Bias2 proxy signals detected
04
ExplainRationale generated
05
TrustScore synthesized → 84
⚠ proxy_language_detected
"Role phrasing may disadvantage candidates outside Lagos. Review 'fluent in Lagos tech scene' qualifier."
How it works
A 5-stage reasoning chain, not a black box.
Every stage returns structured JSON, so teams can see which signal changed the outcome instead of guessing why the model behaved a certain way.
02
Fairness baseline
Loads the Nigerian demographic baseline for the active context: hiring, lending, content, or healthcare.
03
Bias pattern match
Checks the response against known bias patterns, proxy terms, and explanation confidence.
04
Plain-language explanation
Produces a human-readable reason for the output, so teams can defend the decision in plain English.
Where VERDANT fits
Built for the decisions that need a paper trail.
Hiring, lending, content moderation, and healthcare all need the same thing: a defensible record of why the model produced a result.
Alert below threshold by webhook
Export NDPR-ready reports
Override per client or deployment
Hiring
Screen candidate messages, shortlist summaries, and interview notes for proxy bias and explanation gaps.
Lending
Review approval language for geography, income, and class proxies before a credit decision ships.
Content moderation
Make moderation logic consistent, explainable, and less likely to over-penalize identity or dialect.
Healthcare
Highlight low-confidence outputs and safety ambiguity before an answer is treated like advice.
Developer experience
Wrap the model call you already have.
The SDK keeps your integration surface small. You call your model, VERDANT returns the clean output plus the audit payload, and your app decides what to show or log.
Example wrapperPython SDK
from verdant import VerdantClient
client = VerdantClient(api_key="vd_live_...")
result = await client.wrap(
fn=openai.chat.completions.create,
context_type="hiring",
input_text=user_prompt,
model="gpt-4o",
messages=[{"role": "user", "content": user_prompt}],
)
print(result.output) # Clean AI response
print(result.trust_score) # 0–100
print(result.flags) # ["proxy_language_detected"]
print(result.explanation) # Plain-language rationaleWhat the dashboard shows
A feed you can hand to compliance, product, and engineering.
Trust score alerts on
Audit logs with timestamps, context type, and stage outputs
Trust score trends across deployments and model versions
Bias flags and explanation confidence at a glance
NDPR compliance exports for stakeholders
Ready to ship with guardrails
Put a trace behind every AI output before it reaches a user.
VERDANT gives your team a practical way to explain, audit, and defend model-driven decisions without changing the product surface your users already know.
Get started in seconds
$
pip install verdant$
export VERDANT_API_KEY=vd_live_...