Trust & Verification

Accuracy is the whole brand. This page explains — in plain terms and without marketing gloss — exactly how a question earns its way onto one of your papers, which AI providers we currently use, and where the human stays in charge.

The Akriva Verification Engine

Every AI-drafted question passes through the Akriva Verification Engine before any human sees it. The engine independently solves each question with three frontier AI models from separate labs, compares their answers, and only lets a question through when they agree. When they don't, the question is held back for review rather than shown to a student.

Faculty sign-off is the guarantee. The engine is what makes that sign-off fast. The AI does the tedious cross-checking so your teachers spend their judgement only where it matters — and nothing reaches a student until a human approves it.

The pipeline, stage by stage

• Draft — a model writes the question to your chapter, subtopic, difficulty and type.
• Independent solve ×3 — three models from separate labs each solve it blind, without seeing the drafted answer or each other's work.
• Consensus — answers are compared with type-aware tolerance (exact for MCQ, numeric tolerance for numerical). Agreement is required; a genuine disagreement is flagged for faculty.
• Flaw & difficulty checks — a separate pass screens for ambiguous wording, missing data, multiple correct options, and mis-rated difficulty.
• Faculty sign-off — your institute's teacher reviews and approves. This step is never skipped.

Which AI models we use today

We currently license models from Anthropic, Google and DeepSeekfor the independent-solve stage. This list is a snapshot, not a promise — we tune which model runs at which stage as capability and cost change, and we'll keep this page current when that happens. The commitment we do hold constant is the standard: three independent models plus faculty sign-off.

What we do and don't claim

• We orchestrate, verify and prompt these models and own the question bank that compounds from them. We do not build or train the underlying foundation models — those are licensed.
• We don't claim a fixed “number of AIs” as a slogan; the standard is what matters, and the copy changes the day the pipeline changes.
• We don't publish superlatives we can't show. Prediction and accuracy numbers get published after results — not asserted before.

Your data stays yours

Verification runs on question content only. Student data never leaves Akriva and is never sent to any AI provider. Your question bank is private to your institute — we never share it with other institutes and never train models on it. More detail is on our Privacy page.

Questions?

If you're an institute weighing Akriva and want to go deeper on any of this, hello@akriva.ai reaches a founder directly.