Every detail.
Within reach.

Import an exam from PDF or Word, maths included.

• Equations become editable LaTeX, and sub-parts (a)(b)(c) are split out.
• Marks are read from the paper (“5 pts each”), and bonus sections are kept.
• Long papers are processed page by page.

Graphs are cut out of the scan, cleaned up and attached to the right question. They are never redrawn by AI.

AI-written rubrics on “Mark as ready”.

• Kinds: criteria for written answers, calculus tiers (100/80/40/0) for maths, and answer keys for objective questions.
• Maths references: every maths question gets a worked reference solution.
• Spot check: a random 5% of AI-written rubrics must be reviewed before the first session.
• Your edits stay: the AI never overwrites them.

Multiple choice, multiple select, true/false, matching, ordering, fill-in-the-blank, cloze, hotspot (click on an image), calculated (different numbers for each student), worked maths, code, short answer, essay, and case studies with a shared scenario and figures. Sketch tools also support visual maths responses.

A graph-paper canvas gives students tools to show their thinking visually.

• curve tools: parabola, hyperbola with asymptotes, sine, exponential, logarithm, |x|;
• open and closed points, shading and labels;
• adjustable axes.
• The AI grades the drawing itself.

Organised by sections and topics, searchable. You can build an assessment by picking questions or drawing N at random per section.

A visual equation editor for instructors and a 12-category maths keyboard for students.

Formula sheet (PDF) per assessment, viewable inside the locked exam.

Rosters, duplicate assessments or whole workspaces, a recycle bin with 6-month restore, and a preview exactly as students see it.

Assessment settings.

• Time limits from minutes to days, shuffling, back-navigation control, auto-submit.
• Multiple attempts graded by highest, average or latest.
• Result visibility and download.

Native lockdown apps for Windows, macOS (signed and notarized, for Apple Silicon and Intel), and Linux, with automatic updates.

It won’t start inside a virtual machine, a remote desktop session, with a second display, a remote-control app or an active screen share, or (on macOS) with full-screen apps open.

• Full-screen kiosk with shortcuts blocked and the clipboard cleared.
• Screenshots and screen recording blocked.
• Prohibited apps recognised by their program file, so renaming one doesn’t get around it. AI or remote apps opened mid-exam are flagged.
• Debugger launches are refused.

Other apps are hidden during the exam and Mission Control attempts are flagged.

• Answers save within a second.
• After a crash, forced quit or power loss, the student rejoins the same attempt.
• The exam timer is set by the server and keeps running offline.

One-code join, waiting room, maths keyboard, sketch pad, formula sheet, attempt counter.

Only listed students can join; IDs are matched regardless of capitals or spaces.

Changes appear instantly, and Smart proctoring streams from every student every 2 seconds.

Allow or keep blocked, straight from the monitor.

A ranked list with reasons: idle, behind pace, flagged, disconnected, proctoring level.

Expected vs. actual progress per student, a projected finish time, an on-track count, and a class pace chart against an on-pace line.

Per question: how many reached it, median time, live % correct for auto-graded questions, and who is on it now. Callouts highlight the bottleneck, slowest, hardest and most-changed questions.

Integrity over time with live alerts.

Three views (cards, table, and a student × question map), plus filters and search.

Who hasn’t joined yet.

Add time to a running exam (students’ clocks update instantly), auto-open and auto-close, waiting room.

Load-tested with 200 students × 60 questions.

Each student builds a private profile from face geometry, voice, typing rhythm and their usual setup. It’s stored as numbers only, never photos or recordings. It follows the student from exam to exam, course to course and instructor to instructor, gets stronger with each approved exam, and is deleted 12 months after their last exam.

If the face or voice doesn’t match the profile, the exam pauses before the student can answer. The student gets two re-checks; the instructor can allow or keep the block live from the monitor. Every decision is logged.

The face is re-checked every few minutes, so someone swapping in after the start is caught.

A Bluetooth scan recognises smart glasses (Meta and Snap), wireless earbuds (AirPods) and smartwatches nearby. It records device types only, never names, and never pairs. These are combined with what the camera sees:

• smart glasses plus glasses being worn;
• earbuds plus a hand raised to the ear;
• a watch plus repeated wrist glances.
• Bluetooth alone never counts against a student.

Glasses and sunglasses detection, running on the student’s own computer, including “eyes hidden”.

Phones, laptops, monitors/TVs, books or notes, and a second person on camera.

Hand to the ear, glances at the wrist, hands out of view.

Looking away, repeated looks down (a phone in the lap), turning away, and the face leaving the frame. A blink check defeats a photo held up to the camera.

Talking, back-and-forth conversation, and the “voice then answer” pattern (someone dictating). Short audio clips are kept only when things escalate.

Two screens, screen sharing, remote help.

• Blocked at the start: a second monitor, an active screen share (e.g. Zoom), remote-control apps, remote desktop sessions and virtual machines.
• Flagged mid-exam: a screen plugged in during the exam.

Auto-typer rhythm, injected keystrokes, text pasted from outside the exam, and “copy, pause, burst” (typing out an answer from elsewhere).

Each student is compared with themselves. After three approved exams, a student’s normal habits count less against them.

Escalation needs at least two independent kinds of evidence.

Reports are signed and hash-chained on the device. The server replays and re-scores each report rather than trusting the app’s verdict.

Forensic report and evidence bundle.

• A one-clock timeline with lanes for gaze, voice, objects, hands and exam events.
• Key moments, each with possible innocent explanations.
• Captures only when it escalates: stills at amber, short clips at red.
• A downloadable evidence ZIP.

Same internet connection plus near-identical answers, switching connection mid-exam, logging in twice, sharing a computer with another student, and rewriting an answer right after leaving the exam window. Shared lab/classroom connections are recognised automatically and not counted.

A measure-only pilot for rollout, and a K-12/minors mode that stores no images, audio or video at all.

Camera and mic, position, gaze range, blink, typing sample, quiet moment and a read-aloud sentence, with consent built in.

“Followed another student.” Flags a student who lands on the same distinctive answers shortly after a classmate, question after question.

A statistical test for two students sharing unusually unlikely wrong answers.

Genuine answer changes made within seconds of each other on several questions.

Essays, code and worked maths are compared, including “one answer contains the other”. Standard working that much of the class shares is left out automatically, and shared phrases are highlighted side by side.

Integrity report per student and per class.

• Ten indicators roll up into one standing per student: clear, review, high or blocked.
• Plain-English findings, a “when it happened” timeline, and CSV export.

Each flagged student or group is marked “fine” or “concern” with a note, and grades can’t be approved until that’s done.

Two models grade each answer independently. One provides the official proposed grade and the other a second opinion. Disagreements are shown to the instructor and never resolved silently.

• Instructor-aligned co-grading is shaped by co-authored rubrics, grading rationale, and corrections, with alignment checked through instructor-led review.

Equivalent answers are proven correct instantly. A pre-solved reference makes partial credit consistent. Showing work is expected by default, and calculus tiers apply.

Graded by the rubric, criterion by criterion, with the breakdown visible to the instructor.

Sketches graded from the drawing: shape, intercepts, asymptotes, end behaviour.

Identical answers always get identical grades.

Grades in seconds after a student submits.

Edit a rubric and the whole question is re-graded. Change the points and existing grades rescale. Grades you set by hand are never overwritten.

Grades held until you approve them.

How many students could still change letter grade or pass/fail, and about how many minutes of review are left.

Both AI grades side by side with a per-criterion comparison. Choose the AI’s grade, the second opinion, or your own.

Grade a few random, high and low assessments without seeing the AI’s marks, then see how the AI compares and whether it leans harsher or more lenient than you.

• Blind calibration is part of Second Look. The instructor marks a sample without seeing the AI score, then compares the results and rationale.
• The aim is a personally calibrated co-grader that improves alignment over time and across courses.

Questions where the class average is under 50%: read every answer, fix the rubric, re-grade.

Integrity verification in the same flow.

Approving without a full review requires a written reason, which is logged. Exports are marked DRAFT until approval.

• One row per student with a square per question.
• Filters (needs grading, below 60%, integrity, ran out of time), search, and a heatmap matrix.
• Half-screen panels for a student or a question, with inline grading and keyboard navigation.

• Student report: where marks were lost and what to work on.
• Answer sheet: the full record.
• Certificate and class roster.

“What stands out.” Prioritised, plain-English insights, each linking to the detail.

A difficulty-vs-discrimination map, test reliability (Cronbach’s α), and score spread per question.

Learning gaps measured from the rubric criteria students missed most.

score. Scatter plot with correlation and a pacing curve.

Distribution histogram, grade bands, pass rate, and performance by question type.

Model agreement, your decisions, and blind-grading agreement.

• Questions where signals clustered.
• Scores by attention level, and a timeline.
• An identity summary and which exam-app versions students used.

• Profiles are numbers only, with consent, kept 12 months; captures are kept 180 days.
• Internet addresses are stored only as one-way hashes, and Bluetooth records device types only.

Instructors see only their own students. Evidence is served through short-lived access links.

Encrypted nightly off-site backups, snapshots every 60 seconds during live exams, and automatic recovery from database damage.

• Import from QTI, Moodle, Blackboard and Canvas files, plus JSON/CSV into the question bank.
• Export to QTI 1.2/2.1, Moodle XML, GIFT, Blackboard and Canvas.

Grade CSVs for Canvas, Blackboard, Moodle and D2L. An Excel workbook, “export exactly this view”, all answer sheets as one PDF, and student reports as a ZIP.

Google, Microsoft, LinkedIn, Apple or your institution’s SSO.