What ControlAI Can Do¶

ControlAI isn't a single feature — it's a whole set of capabilities built around the class session. This is an overview catalog: what the system sees, calculates, and shows, grouped by meaning.

Some of it works today; some is on the roadmap (marked separately and collected in Roadmap). What each number actually means is covered in Metrics and What They Mean. Who sees what — in Inside the Admin Panel.

1. Whether the class actually happened¶

The first and simplest question ControlAI answers for every class, every day is whether the session took place at all, and to what extent:

Class not held — the system sees from the recording that nothing happened in the room (no speech, empty room) and records this as a fact, not as someone's complaint.
Punctuality — whether the class started and ended on time (how many minutes late it started, how many minutes early the instructor let the group go) — down to the minute, straight from the recording, not from a log.
All of this is objective and requires no inspectors: not a spot check by a floor monitor, but full coverage of every class in every department.

Checking whether a class was held and its timing works in a room of any size — from a small seminar room to a large lecture hall.

This also protects the conscientious instructor: if you held the full class, it's visible and on record — subjective complaints like "you weren't there" no longer hold up.

2. Speech analysis of the class¶

Talk-time balance — how long the instructor spoke versus the students.
Questions — how many questions the instructor asked.
Pauses and silence — how much of the class passed in silence, and how long the longest pause was; whether the class "sagged."
Share of class language — what portion of the class was conducted in the group's language of instruction, and what portion in another language. The language of instruction (Uzbek / Russian / mixed) is set when the group is created — reports arrive in the group's language, and speech recognition handles the Uzbek–Russian code-switching typical of these classrooms with confidence.
Number of distinct speakers — an estimate of how many different voices were heard in the class: was it a monologue or a dialogue with the group.
(roadmap) interaction pace (how many instructor↔student exchanges per 10 minutes), how speech is distributed among students ("3 students accounted for 90% of talk time"), language mix over the course of the class, audio-quality monitoring in the room (catches a "dying" microphone before report quality drops), plus offensive or inappropriate language, praise and encouragement of students, calling students by name in seminars and practical classes (not a meaningful signal in a hundred-person lecture stream), and drifting off-topic.

Speech metrics work for any type of class — lecture (ma'ruza), seminar, practical (amaliy), lab; the "good pattern" simply looks different for a lecture than for a seminar. Transcript completeness in large lecture halls is verified with an audio audit at setup (an external microphone is added to the camera's audio input where needed).

3. Video analysis of the class¶

Empty room — the system detects when a class effectively didn't happen (cross-checked against section 1).
(phase 2 — roadmap) attendance (davomat) — how many students were in the room versus how many were expected; latecomers — how many students arrived after the class started, and by how many minutes; further down the roadmap — who's on their phone, who's active versus passive for the whole class, heads down on the desk (signs of sleep, fatigue, or boredom).

To be candid: today video is used to confirm that a class took place; automatic attendance from video is phase 2, not a current feature.

4. "Lesson DNA"¶

A visual color timeline of the class: how the 80 minutes broke down — where the instructor spoke, where students spoke, where there were questions and discussion, where there was group work, and where there was silence. One glance shows the "shape" of the class: did the instructor lecture in monologue the whole time, or engage the students. Alongside it, the system flags key moments of the class (both good and problematic) tied to a timestamp — for example, "14:18 — 5 minutes of group work" or "14:27 — 6 minutes of silence."

5. Methodology compliance: the "open lesson" — now for every class¶

This is the "quality inspector" built into ControlAI — it checks the class against your own university's standards. It's the deepest part of the product.

Today, an open lesson (ochiq dars) with a review commission happens once or twice a year for an instructor — and everyone prepares for it. ControlAI fills out your university's own lesson observation form for every single class — the open lesson stops being a special event and becomes the norm.

How it's configured: the university uploads its own lesson observation form (the one the commission currently uses), its own methodology, and its syllabus / work program. From then on, ControlAI evaluates every class against exactly these materials — the criteria aren't hard-coded into the system, they're yours.

What gets checked (a typical form has 12–18 criteria across several blocks; your university's will be its own):

Class organization — order in the room, contact with students, on-time start and finish.
Class structure — whether the class followed its stages (warm-up/review → presenting material → practice → independent work) and whether each stage got adequate time. (part of the stage-level checks is on the roadmap)
Teaching technique — whether the instructor draws answers out of students rather than just telling them, asks comprehension-check questions, gives clear assignments and verifies they were understood, monitors student work, how feedback is given and mistakes reviewed, and whether the material is pitched to the group's level.

What comes out for every class:

A filled-in observation form — a rating for each criterion, a short evidence quote from the class, and a note on how confident the AI is; disputed items are honestly flagged as "needs a human look."
An overall methodology score (1–10) — not a bare number, but broken down by criterion with a confidence level. This is internal analytics for decisions and development, not a substitute for official certification; at the university's discretion it can serve as supporting evidence for internal KPIs.
Topic coverage — what's already been covered from the syllabus, what's partially covered, and what hasn't been touched yet (cumulative per topic).
Human review — the university chooses the mode: mandatory review by the quality-control office (ichki ta'lim sifati nazorati bo'limi) or the head of department before publishing, or auto-publish; either way, the reviewer's corrections make the system more accurate.

The AI flags disputed and subjective items as low-confidence, so an incorrect assessment of an instructor's work is never presented as fact.

6. Instructor analytics¶

Class-level metrics roll up into a profile for every instructor:

Rating — an overall performance score (how it's calculated — in How the Instructor Rating Is Calculated);
What makes up the rating — a breakdown by factor (talk-time balance, questions, language, punctuality) with clear explanations;
Skills radar — strengths and weaknesses across several axes;
Key moments from classes — good and problematic episodes tied to a timestamp;
8-week trend — whether the instructor is improving or slipping (movement is visible within the semester itself);
"Needs attention" feed — the system surfaces what matters on its own: a class that didn't happen, a sharp drop in rating, too much silence, a camera problem, an outstanding result.

For the instructor, this isn't "a camera watching from above" — it's a personal dashboard: their own stats, their own trend, and a fair, consistent evaluation instead of occasional subjective spot checks.

7. Coaching and growth¶

Coaching buckets — the system sorts a department's instructors into "recognize," "support," and "watch," and suggests a concrete next step.
Goals — an instructor, a head of department, or a mentor for junior instructors sets a goal (for example, "increase student talk share in seminars"), and progress is tracked automatically.
Reviews/evaluations — the head of department or the quality office records a review of an instructor's work; those edits also accumulate and improve the system.

8. Risk of a group "falling behind"¶

ControlAI flags in advance which groups are at risk of falling apart: it computes a per-group risk score (0–100), tracks engagement trends from speech signals (and, from phase 2, attendance from video), names the likely cause, and recommends an action — so the department can step in before absences, academic debt, and expulsions pile up.

9. Comparative analytics: departments, faculties, years, courses¶

A university is a structure, and ControlAI compares its levels against each other on key indicators (talk-time balance, questions, punctuality, share of classes held) — this already works today, on live semester data:

a university-wide summary for the rectorate — comparing faculties against each other on one screen;
department vs. department and an instructor rating within a department;
breakdowns by year of study (bachelor's and master's) and course vs. course — for example, how "Higher Mathematics" is taught across different faculties, or how "Transport Operations" goes for different instructors;
semester trends — top performers and laggards, who's improving and who's slipping.

Planned — end-of-semester summary reports (semester wrap-ups by instructor, department, and faculty, and comparisons between semesters): see Roadmap.

10. Reports and summaries¶

Class report — metrics plus a short, clear write-up of how the class went, in the group's language of instruction (Uzbek or Russian).
Daily summary for the dean — every class in the faculty for the day in one list, problems surfaced first: what didn't happen, where there were late starts, where the warning signs are; the rectorate sees the same kind of summary for the whole university.
Personal digest for the instructor — their own stats and a short recap of the class (framed as "help," not "a grade from above").
History of all classes — any past class can be found by filter (group, instructor, room, department, date) and its results opened.
(roadmap) a weekly summary with trends across instructors and groups.

11. Administration and operations¶

Around the analytics is everything needed to actually run the system at a university:

Class schedule — the semester's grid of classes (who, when, which group, which room, what type of session); this is exactly what tells the system when to go fetch a recording. The schedule is entered and maintained in the panel; HEMIS integration is on the roadmap (phase 3). ControlAI warns about conflicts (an instructor or a room double-booked).
Courses and groups — a catalog of course subjects; each group has a year of study and degree level (bachelor's or master's), a field of study, and a language of instruction.
People and groups — instructors, groups, students; invitations; consent status.
Rooms and equipment — camera status (ready / needs a microphone / offline), connectivity checks. Universities typically already have cameras and a video recorder — ControlAI connects to the existing infrastructure and installs nothing in the classrooms (aside from an external microphone where the built-in one fails the audio test).
Billing — volume is measured in hours of analyzed classes (terms and volume per contract); the panel shows daily usage, billing runs on actual processed minutes of class time, and a class that didn't happen isn't included in the billed volume.
Workload tracking — hours based on classes actually held (not on the plan) — an objective basis for confirming teaching-load fulfillment; exports for accounting and the academic affairs office.
Operations — a task list (camera offline, no audio, processing failure, class not held, consent pending) and a coverage summary: how many of the scheduled classes were successfully processed on time.

12. Breakdowns and comparisons¶

A dedicated section for digging into the numbers: pick a department or course subject, a metric, and a period, and pull up a breakdown — for example, comparing instructors in the "Rolling Stock" department by number of questions over a month — and export the result.

The key takeaway from this chapter: for every class, ControlAI answers the question "did the class actually happen?", analyzes the speech and teaching methodology (against your university's own observation form — an "open lesson" for every class), builds instructor, department, and faculty analytics on top of that, plus coaching and early warning signals for groups, prepares reports in the group's language of instruction, and around all of it provides everything needed to run things — schedule, courses, people, equipment, billing, workload tracking, and quality control. Video-based attendance and some of the more advanced signals are on the roadmap, and they're marked honestly as such.

→ Next: 05. Metrics and What They Mean — we'll break down every number in plain language and show what a good class looks like.