Abhyas Studio

Chapter 1 · The problem

The mentors who carry India's reading and math reforms into classrooms have nowhere to practice their job

Start with a girl in the second row of a Class 2 room in a rural school. She is working through 42 minus 17, and writes 35: she took the smaller digit from the larger in each column. She is not unusual: in rural India, only 27 of every 100 children in Grade 3 can read a text meant for Grade 2.

Every state now runs a program to fix this. The plans are made in state capitals, and they reach the girl through one group of people: the block and cluster mentors who visit schools, watch lessons and advise teachers. They are the main layer of the system that regularly walks into classrooms.

Picture one of them. Call him Raghu sir, a composite I built from what real mentors told me. He is in his early fifties, with 26 years in schools. He runs his own school and is responsible for twelve others. On a visit day he watches Lakshmi garu teach her Class 2 math period. Afterwards they sit under the neem tree in the yard, and he has about ten minutes. He tells her what went wrong: the counting sticks stayed in their tin, the children who need the most help were left behind, the workbook is not finished. She nods and writes it down. He writes one line in the school's register and drives to the next school.

Nothing changes, and nobody ever showed him another way.

A mentor's job has three parts, and none of them is practiced

Coaching raises learning only when the visit is about teaching

This matters because coaching teachers in their own classrooms is one of the best-proven ways to raise early learning. When the coach does not know how to coach, visits drift into inspection, and inspection changes nothing.

So the people meant to support teachers have never practiced their own job, and there is nowhere to practice it. A human trainer cannot sit beside thousands of mentors before every visit. Pilots learn to land in a simulator, where a mistake costs nothing. Mentors had nothing like it.

Sources: ASER 2024 (Rural), January 2025; the studies listed under each exhibit; Raghu sir is a fictional composite from my interviews, and every event around him comes from those interviews.

Chapter 2 · Who it is for

I designed for two kinds of mentor who fail in opposite directions, and for the phones they carry

I read the state's own job charts and visit protocol before designing anything. Then I interviewed eight mentors, five Complex HMs and three MEOs, about their last real visits, before they saw anything I had built. None of the eight had ever been shown how to give a teacher feedback. Two portraits came out of the interviews, and they get feedback wrong in opposite ways.

The same few moves fix both mistakes. That is why every practice room trains five moves at a time, with feedback after each try, rather than a long syllabus.

Each fact about their lives became a design rule

A link on WhatsApp, not an app from a store

Mentors already live on WhatsApp, and the state already sends them links there. So Abhyas reaches them as a link that installs to the home screen, with no store account and no approval queue. A fix made in the morning reached every tester's phone on their next open.

Sources: Interviews and first study, 26 Oct 2026; persona and design-rule documents in the project repository.

Chapter 3 · Design

Every design rule traces to research on how adults learn a skill

You can read about a good feedback conversation and still lecture when the moment comes. Medical students practice on actors playing patients for the same reason. I wrote an evidence review before designing, and its strongest findings became the product's rules.

The job became 91 skills, and each room trains five moves

"Be a better mentor" cannot be practiced. "Ask the teacher's view before giving your own" can. I wrote one competency framework for each part of the job, from the state's own job charts and visit protocol: 28 skills for teaching, 36 for mentoring and 27 for leadership. Each skill is described at three levels, from Not yet to Skilled, in terms of what the mentor says or does.

A list of 36 skills is for designers. The mentor hears five moves per room, from the senior mentor:

One rule runs through all three: credit is for what a move achieves, never for a phrase. A mentor cannot earn a point by saying "what do you think?" and then ignoring the answer.

Scenarios are written like real visits, with a decoy and a hidden reason

A good scenario is a fair test of judgment. Everything the mentor needs is visible if they look, and nothing is labeled for them.

Next An outside pedagogy expert reviews the three frameworks and the scenario content next. The review queues are written.

Sources: Evidence review and research notes; three competency frameworks and item banks; scenario scripts. All in the project repository, Oct 2026.

Chapter 4 · What I built

Three practice rooms rehearse the whole job; seven scenarios are live, and mentoring is built deepest

Raghu sir holds his phone sideways and finds himself at the back of Lakshmi garu's Class 2 math lesson, drawn as an animated scene. When he turns the phone, his view of the room turns with it. He notices the counting sticks locked in their tin, and taps to note it.

After the lesson he sits down with Lakshmi garu, an AI character, and gives her feedback by speaking, as he would in a real school. She replies out loud, in Telugu or English. Her body language follows how he treats her: she folds her arms when he lectures, and leans in when he asks what happened.

1. 

   1Watch the class

   Look around the room by moving the phone. Tap to mark a moment worth raising.

2. 

   2Talk to the teacher

   A spoken conversation with Lakshmi garu. She is polite and busy, and she has a reason for what you saw that she will not volunteer.

3. 

   3See it from her side

   A senior mentor, Saraswathi madam, replays the key moments from Lakshmi garu's side: what she was thinking each time she opened or closed.

4. 

   4Get coached, then try again

   One strength, one thing to work on, and words to try, built from your own sentence. Then you say that moment again, better.

Every room follows the same loop: meet the person, do the job, see it from the other side, get coached on one thing, try that moment again, hear how the senior mentor would do it, plan the next real visit, and practice the move in three short drills. It takes about 17 minutes.

In the leadership room the mentor sits at a head teacher's desk, sorts an in-tray, takes a call from the block officer and hands the school over before leaving. Each room is a ladder of three levels, from "start well" to the hard cases experienced mentors meet, and the help on screen falls away as a mentor climbs.

The Class Engine decides what each of 41 children understands

A class of 41 is harder to simulate than one teacher. If an AI model invented the children's answers, a mentor could never trust what the class learned. So a second engine keeps a record for every child, and every chorus, answer and slate is worked out from it.

The AI's job in the classroom is narrow. Code decides which child speaks and what they mean; the model only puts it into words, and a check rejects any line that changes a number or adds a name. The leadership room follows the same principle: code scores every choice, and the AI judge may lower a level, never raise it, and only by quoting the mentor.

Watch the three-minute film

I made this film for the hackathon on 1 November 2026. It follows Raghu sir, a fictional cluster mentor, through all three rooms, and every app scene is a live run of the app. It is narrated, so turn the sound on.

Sources: The app as of 1 Nov 2026; scenario ladders; the hackathon film, 1 Nov 2026.

Chapter 5 · The Trust Engine

The AI teacher opens up only to a mentor who earns it; a plain chatbot opened up to one who only sounded kind

Lakshmi garu keeps the sticks in the tin for a reason, and a good mentor finds it out by asking. A plain chatbot given her description tells the reason to anyone who sounds kind. I tested that before building the engine: it told a mentor who only stacked kind phrases, at turn 4, in 3 of 3 runs.

AI models are agreeable. If kind-sounding phrases work on the simulator, mentors learn to use phrases instead of listening. So in Abhyas, the model writes her words, and code decides how she feels.

Code keeps her openness score; the model only writes her words

The Trust Engine keeps a score from 0 to 100 for how open she is. She starts at 65. After each thing the mentor says, an AI model labels the move (asking, affirming, telling, blaming) and the engine turns the labels into points by fixed rules:

• Trust builds slowly and breaks fast. One turn can add at most 6 points, but judging or threatening her can cost up to 15. If her openness falls below 15, she ends the conversation.
• Praise next to a criticism counts for nothing. A compliment followed by "but you should" is heard as a set-up, so any praise in a turn with a loss earns zero.
• Facts unlock at set levels. What she will reveal is locked behind gates at fixed scores. The model is never given a locked fact, so it cannot leak one, and code checks her reply twice before she speaks.

A curious mentor heard her reason at turn 4; a telling mentor never did

I ran scripted mentors through the real models. Each visit opens with a minute of small talk in the staffroom, which moves her starting score up or down from 65. A curious mentor started at 73 after a friendly minute, climbed every turn, and at turn 4 the last gate opened and she said why the tin stays shut. A telling mentor started at 61, held level, then lost 3 to 4 points a turn, fell below 55 at turn 3, the point where she stops sharing, and never heard it. She kept saying "Okay madam, I will try". The curious mentor's Telugu run reached the same point at the same turn.

Three refinements the probes surfaced

The engine worked as designed in scripted probes, and the same probes pointed to three refinements. Her words lag one turn behind her score. Asking exactly the right question too early can use up its value. And the AI labeler sometimes reads the same meaning differently in Telugu and English. Each has a planned fix. To test next Next, experts mark consented real conversations turn by turn.

Sources: Trust Engine rules in the configuration file; scripted probes on the live models, 25 and 31 Oct 2026; chatbot comparison, 26 Oct 2026; 616 synthetic-user runs.

Chapter 6 · Fair feedback

AI models do the language work, and code makes every decision, down to each point of feedback

When Raghu sir ends the conversation, he needs two answers: how did he do, and what is the one thing to work on next. Two mentors who say the same thing must get the same answer. So the AI judge reads and quotes, and code checks and decides.

The coach cannot invent a weakness the scorer did not find, because it never sees anything else. Its job is tone, not judgment: to say a hard thing the way a respected senior colleague would.

Every AI call goes through one gateway with fixed rules

Every request to a model goes through one gateway in my code. Each call has a fixed shape for its answer (0 of 412 live calls broke it), a length limit, and a cost record in rupees. A visit stops spending at Rs 25. If a call fails, it is retried, then sent to the other Claude model, then to a backup from a second AI company, then replaced by a scripted line. That plan was tested for real on 31 October, when testing hit Claude's usage limit: the live app switched to the backup by itself and ran a full test visit on it.

Testing caught two unfair judgments, and I fixed both

The judge marked a strong but short conversation "Not yet" on skills that never came up; a skill with no evidence now counts as "not seen", and "Not yet" marks on that conversation fell from 9 of 19 to 4 of 19. The drill check had the bigger problem: on 30 October it accepted none of nine natural answers from real mentors, because it judged whether an answer matched the example rather than whether it made the move. The new version accepts 320 of 326 good test answers and rejects all 128 wrong ones. To test next Next, I re-run the nine real answers through it.

Every protection sits outside the models

Two of these choices cost speed, on purpose. Her voice could start sooner if she spoke while her reply was still being written. I refused, because she would say words nobody had checked.

Sources: Gateway, scorer, coach and safety checks in the project repository; live runs log, Oct 2026; fallback test, 31 Oct 2026; drill test set of 454 answers.

Chapter 7 · Testing

Four rounds with different users each changed the product before the next round began

Mentors' time is precious, and a broken app wastes it. So software played mentors first. Then eight mentors, three experts and finally two working mentors in the field each tested a newer build, and each round's fixes shipped within a day.

Eight mentors took it seriously and named changes

After a first, partial practice, 7 of 8 named a change for their next real visit, and the changes were small and real: ask one child by name, look at a struggling child's workbook, sit longer, talk to the teacher before class. They rated realism a median 3.5 of 5 and said what would raise it: the teacher was "too neat, too cooperative". 8 mentors

In the field, every spoken turn worked, and the review ran long

21 of 21

spoken turns in the feedback conversation went through

111 of 115

voice turns across the whole visit worked

35 min

for one mentor's full visit, against 17 designed

The conversation, the heart of the product, worked, and nobody showed impatience. The extra time went almost entirely to the review screens. A facilitator named the cause: too much information on each screen, so mentors needed help to move through the review. That set the target for the redesign below.

The MEO went back through the cases after the session, and a facilitator relayed what he said:

He also said who should get it first: "If this app is given to newly joined CHMs, or to CHMs and MEOs without much experience, it will really help them a lot" (relayed, and translated from Telugu). That advice now shapes the pilot.

Behavior data, not anyone's words, drove most fixes

The testers had not agreed to share what they said, and practice is private. So the app records behavior instead: 18 kinds of event, such as which screen was open and for how long, every tap by the button's internal name, replays, idle time, and the timing of every step of every spoken turn. A dashboard reads it against a target for each metric.

A measured redesign of every screen

After the field round I stopped guessing. An automated walk captured all 210 screen states at the exact screen sizes testers used, and scored each one from 0 to 12 on six usability tests, such as words on screen, number of choices and buttons smaller than a fingertip.

5.0 → 1.0

average screen load, on a 0 to 12 scale (lower is lighter)

35 → 24

stops in one visit

51 → 30

taps in one visit, approximately

The rules that came out of it are simple: the voice carries the content, one main button per screen, everything fits a phone held sideways, and one choice at a time. Being built The redesign was approved on 1 November and is scored from its designs; the next field round will test it against the same targets.

Sources: Synthetic-user grid (616 runs); study kit and first study, 26 Oct; expert feedback, 29 Oct; field findings, 30 Oct; the facilitator's-hunch investigation, 31 Oct; screen audit of 210 states, 1 Nov 2026.

Chapter 8 · Results

Targets were set before the first user; the AI's behavior passed, and speed and screens are next

I wrote a target for every important metric into the configuration file before testing, from research where it existed. So every result below is read against a bar set in advance, and each miss has a planned fix.

The first target for the judge could not tell a good judge from a bad one

To trust the judge, you compare it with human experts scoring the same conversations. Agreement is measured with Cohen's kappa, where 0 is chance and 1 is perfect. My first rule said the judge must agree with experts at least as well as they agree with each other, and reach a kappa of at least 0.6.

Before using the rule, I tested it on simulated scores, and it failed. Two experts typically agree at a kappa of about 0.53 on skills like these, so the fixed bar was higher than the experts themselves. A judge exactly as good as an expert would pass only about one time in five. A clearly worse judge would still get at least one skill scored, by luck, about one time in four.

The new rule asks that the judge agree with the experts nearly as well as they agree with each other, within a fixed margin of 0.10. A skill on which experts themselves disagree is coached but not scored. Proving this takes 100 conversations, each scored by two experts, and scores stay marked provisional until it passes.

Sources: Configuration file (evaluation targets, latency, caps); evaluation plan; live runs log; field findings, 30 Oct 2026; simulation of the judge rule, Oct 2026.

Chapter 9 · Cost

A practice visit costs about Rs 14, and weekly practice for a whole state would cost 2% of its visit spend

An app for a mentor on a budget phone in a rural sub-district has to be cheap, work on a weak signal, and not keep him waiting. The first two hold today, and speed is the next fix.

Across six live visits, AI and speech cost Rs 13.76 per practice visit on average, about 16 US cents, against a hard cap of Rs 25. The judge is about two thirds of that, because it reads the whole transcript once for each of its four parts. Storing that shared text so it is read once is the next saving.

To put Rs 14 in context: I estimate one real school visit costs the state about Rs 1,045 in the mentor's time and travel. One practice before a visit adds about 1.3% to that.

The return is simple arithmetic. Rs 34 lakh buys about 3,250 visits, 2% of the 165,000 a year. So if practice turns 1 visit in 51 from no change into a useful one, it breaks even, counting a useful visit as worth what a visit costs.

Streaming speech is the next fix for 6-second replies

In conversation, a pause of two or three seconds feels natural. Today she takes about 6 seconds on a 3G Android phone (5.9 seconds at the median for the field testers). The delay comes from doing things in order: his voice becomes text, the move is labeled, the score updates, her reply is written and checked, and only then spoken. I will not let her reply be written before the score updates, so the next big saving is streaming speech in short pieces both ways, sending only text that has passed the check.

It opens on a weak signal, and practice runs in the cloud

Tested on the live app, 1 Nov On slow 3G, the first load is about 0.6 MB and the Start button appeared in 10.5 seconds. With the network off, the installed app still opens and says clearly that it needs a signal. If the signal drops mid-conversation, his turn is kept and can be sent again, so nothing he said is lost. Practice itself needs a signal, because the teacher, judge and coach run on cloud services.

Sources: Live runs log (six costed visits, 13 timed visits); field findings, 30 Oct 2026; visit cost model; the weak-network test on the live app, 1 Nov 2026.

Chapter 10 · What comes next

Four tests come before a pilot, then four stage gates lead to every state

Testing with real mentors showed that the conversation works and that mentors take practice seriously. The next tests measure whether practice makes them better at the job, starting with four checks before a district pilot.

The same stage puts mentors in the teaching and leadership rooms, sends all 1,255 content items (scenario lines, drills and coaching notes) to a pedagogy expert and a native Telugu reviewer, and runs the attack tests against the live models.

The habit has to be designed, because only 3 of 8 would practice on their own

Only 3 of 8 mentors said they would open the app on their own before a visit. One put it simply: somebody has to tell them, "sir, today at 11 o'clock, sit with me". So I propose a practice slot in the monthly cluster meeting the MEO already runs, and a reminder the evening before a planned visit. Practice stays private and voluntary: 8 of 8 said only they should see their practice, and nobody sees who skipped.

The pilot's main outcome is how mentors give feedback on real visits.

A second state does not need a new app. Everything specific to one state, its people, lessons, language and visit protocol, lives in one content pack, and every number lives in one configuration file. The engines stay the same.

Back to the girl in the second row, and her 35. Bundles of ten sticks are how a Class 2 child sees what borrowing means in a subtraction, and hers stay locked in a tin. Getting them out takes all three parts of Raghu sir's job: knowing how bundles teach borrowing, asking Lakshmi garu why the tin stays shut, and settling the missing kit with the head master at the next review meeting. Abhyas exists so that he has rehearsed each of them before he walks into her room.

Sources: Gaps list, evaluation plan and pilot design, 1 Nov 2026; first study, 26 Oct 2026 (habit and privacy); field findings, 30 Oct 2026.

Reflections

Rules in code made every fix testable, and a screen audit now does the same for design