Apurv Mehra | Portfolio

Most of our work at Disha AI has focused on students — test prep aspirants, coding learners, working professionals. But over the past year, a partnership with Simple Education Foundation (SEF) and BITS Pilani pushed us to flip the lens: what if AI's most impactful role in education isn't tutoring students, but helping teachers get better?

SEF runs a Diploma in Education program that trains individuals to teach K-12 students. Their trainees — many of them first-generation educators — face a problem that no amount of pedagogy coursework can solve: they rarely get consistent, structured support to reflect on what's actually happening in their classrooms.

What went well today? What challenged me? What should I try differently tomorrow?

These are the questions that separate good teachers from great ones. And they're the questions that almost never get asked in large-scale teacher training programs, especially in low-resource settings.

The Problem Up Close

Teacher training in India has a structural gap. Trainees spend most of their curriculum on theoretical learning and only a fraction in actual classrooms. When they do get practical experience, the mentorship infrastructure is thin. Many trainees hesitate to seek feedback from faculty — fear of judgment, lack of confidence, or simply no structured channel to do so.

The result: pre-service teachers learn about teaching but don't develop the reflective habits that make someone genuinely effective in a classroom. In-service teachers face a different version of the same problem — they're teaching daily but have no consistent space to process what's working and what isn't.

This is the gap we set out to fill. Not with another AI tutor that delivers content, but with an AI reflection partner that helps teachers think.

What We Built

The SEF AI Assistant is a Telegram-based self-reflection bot powered by GPT-4. The design philosophy was deliberately restrained — we didn't want it to teach; we wanted it to listen, question, and reflect back.

Here's how it works:

Structured check-in. Each session starts by gathering context: the trainee's background, teaching stage (pre-service or in-service), current challenges, comfort with technology. This creates a persona that calibrates all subsequent interactions.
Reflective conversations. The bot engages trainees in guided discussions about their day — what happened in class, what surprised them, what frustrated them. It asks open-ended questions for exploration and closed-ended questions for clarity.
Root-cause orientation. Rather than handing out solutions, the bot nudges trainees to identify the root of their challenges. It surfaces relevant teaching principles and asks how they might apply. This is a deliberate design choice — we wanted trainees to build problem-solving capacity, not develop dependency.
Session summaries. At the end of each conversation, the bot provides a structured summary: current progress, new learnings, strengths, and areas to work on.

We deployed on Telegram because it's low-bandwidth, familiar, and accessible on basic smartphones — critical for our user base.

What Users Actually Did With It

We ran two iterations — August 2024 and December 2024 — collecting feedback through surveys and focused group discussions across 21+ trainees.

The usage patterns were revealing. Some trainees used the bot 4-5 times a week for 30-45 minute sessions. Others checked in once a week for 15 minutes. The variance told us something important: reflection isn't one-size-fits-all. Some people need a daily debrief; others need a weekly check-in. The bot accommodated both.

What worked well:

Bilingual support. Trainees could converse in Hindi or English, which dramatically widened accessibility. Several users explicitly called this out as a key factor in their comfort with the tool.
Activity suggestions for classrooms. The bot could suggest interactive teaching methods — using stories, experiments, movies — that trainees then brought into their classrooms. Subject-specific suggestions (like teaching transpiration through hands-on experiments) were particularly valued.
Lesson planning and classroom management. Trainees reported that structured reflection with the bot directly improved how they planned future lessons and managed classroom dynamics.
Special needs support. One group specifically highlighted the bot's ability to suggest strategies for students with learning difficulties — an area where trainees felt most underprepared.
Root-cause reflection over direct answers. This was perhaps the most important signal. The bot didn't just hand out advice; it helped teachers reflect on their problems and arrive at their own solutions. Multiple trainees noted this as more valuable than getting a direct answer, because the solutions they discovered themselves were more applicable to their specific contexts.

What didn't work:

Response latency. The bot was slow — sometimes 20 seconds for a reply. We improved this in v2 by streaming responses in chunks and adding a typing indicator, but it remained a friction point.
Formatting issues. Complex responses lost their structure on Telegram — bullet points misaligned, punctuation dropped. A small thing, but it eroded trust in the bot's competence.
Repetitive questions. After extended use, some trainees noticed the bot cycling through similar prompts. Session memory and conversational variety need significant improvement.

The Insights That Matter

1. Reflection is the missing layer in teacher development

We've spent two years building AI tools for learners — test prep students, coding learners, professionals building communication skills. Every time, we've found that the bottleneck isn't information or even practice. It's the capacity to reflect, adapt, and grow from experience.

For teachers, this is even more acute. A teacher who doesn't reflect after a tough class will repeat the same mistakes. A teacher who does will compound improvements over weeks and months. The SEF bot gave trainees a structured, judgment-free space to do exactly that — and the ones who used it consistently reported tangible improvements in their teaching.

2. AI works best when it refuses to give direct answers

This echoes what we found with Sakshm AI, our Socratic AI tutor for programming. In that deployment, students who engaged most deeply with guided questioning actually reduced their reliance on AI over time — scaffolding transferred capability rather than creating dependency.

The same pattern showed up here. The bot's refusal to just hand out "5 tips for classroom management" forced trainees to think through their specific situations. Multiple FGD participants said this approach was more useful than a direct answer because challenges vary case by case. The best AI tools don't answer questions. They help people find their own answers.

3. The trust dynamic is different for teachers vs. students

With students, we've repeatedly found that AI lacks the authority to push. Students compare AI feedback to ChatGPT and dismiss it. They need a human face behind the tool before they'll take it seriously.

With teachers, the dynamic was subtly different. These trainees weren't looking for an authority figure — they were looking for a safe space. Many hesitated to seek feedback from mentors for fear of judgment. The AI bot, precisely because it wasn't human, became the safe channel. It was the place where you could admit "I completely lost control of my class today" without worrying about consequences.

This is a genuinely different use case for AI in education. Not AI as authority (which doesn't work), but AI as a judgment-free mirror.

4. The human-in-the-loop isn't optional — but the loop is different here

In our UPSC work, the lesson was blunt: AI can't replace teachers. Students need a human anchor for trust and accountability. In our Comuniqa research, the hybrid group (AI + human expert) consistently outperformed either alone.

The SEF project adds nuance. The bot isn't replacing SEF's trainers or coaches. It's extending their reach into the hours between sessions, the evenings after a tough class, the moments when a trainee needs to process but doesn't have a mentor available. The human-in-the-loop here is the training program itself — the bot is effective because it operates within a broader human system of curriculum, cohorts, and mentorship.

5. Platform matters more than you think

Trainees asked us to move to WhatsApp or build a dedicated app. Telegram worked technically but wasn't where their daily communication happened. In low-resource, low-tech-familiarity contexts, meeting users on the platform they already use isn't a nice-to-have — it's the difference between adoption and abandonment. We learned this same lesson with BlendNet, where local retailers as intermediaries drove adoption because they were already embedded in users' daily lives.

What's Next

The bot is functional and genuinely useful, but it's v1. The next iteration needs to address:

Faster responses. Latency is a trust-killer.
Better conversational memory. The bot needs to track growth over weeks and months, not just within a single session.
Structured output formatting. Crisp, readable responses matter — especially on mobile.
Platform migration. WhatsApp or a lightweight dedicated app, based on where SEF trainees actually spend their time.
Longitudinal impact measurement. We know trainees liked the bot. We don't yet know if the trainees who reflected consistently became measurably better teachers.

The Bigger Picture

Across everything we've built — AI tutors for test prep, Socratic coding assistants, communication coaches, professional skilling tools — a consistent pattern has emerged. AI's role in education isn't to deliver knowledge. It's to create the conditions for reflection, practice, and growth that would otherwise require a human being who isn't available at that moment.

For teachers specifically, the implications are profound. India needs millions of better-trained teachers. The infrastructure to mentor and support them doesn't scale. An AI reflection partner won't replace the mentor. But it might be the thing that helps a first-generation educator in a rural D.El.Ed program develop the reflective habits that turn a trained teacher into an effective one.

As I wrote earlier, the real bottleneck in Indian education isn't skill — it's agency. The capacity to reflect on your own practice, identify what needs to change, and act on it. That's exactly what this bot is trying to build. Not for students this time, but for the people who teach them.

This project was built by Disha AI in partnership with BITS Pilani and Simple Education Foundation. Related reading: Authority, Not Accuracy, Is What Makes Feedback Work and Learnings from Building an AI Tutor for India's Toughest Exam.