Yukti: Conversational AI voice agent mentorship platform

The problem

Engineering students across India face uneven access to mentorship, industry exposure, and career opportunities depending on which institution they attend. Students at colleges outside the major metropolitan centres often have fewer connections to employers and fewer chances to build the kind of practical project portfolio that strengthens a job application. Many find themselves caught in a cycle: without project experience, it is hard to secure internships, and without internships, it is hard to build the experience that employers look for.

The team at Sustainable Living Lab India (SL2) had been working on upskilling programmes and had developed a simulated internship platform to give students access to real industry problem statements. However, they found that access to problem statements alone was not enough. Students also needed hands-on mentorship to guide them through building AI and machine learning projects.

In government-delivered training programmes, mentor-to-student ratios can be as high as 1:60. Even with corporate social responsibility support from technology companies, maintaining a ratio closer to 1:10 or 1:15 is not practical at scale.

At the same time, SL2 observed that when students used existing AI tools for support, many were putting queries into the tool and accepting the output without engaging deeply with the problem. The team wanted to find an approach where AI could support learning rather than bypass it. Language was also a critical factor. Many students are more comfortable working in Tamil than in English, but technical resources and job interviews operate predominantly in English. SL2 wanted a tool that could meet students in the language they were comfortable with while also supporting their skill development and confidence in English.

The solution

Sustainable Living Lab India built Yukti, a bilingual voice AI mentor designed to help students work through the process of building AI and machine learning projects for their portfolios. Yukti is currently being piloted within the Tamil Nadu Skill Development Corporation (TNSDC) and operates in both Tamil and English. Voice was the preferred channel because spoken, conversational mentoring is more accessible to students, and it keeps learners engaged in a back-and-forth dialogue that more closely mirrors guidance from a human mentor.

Rather than providing direct answers, Yukti engages students in conversation, asks follow-up questions, and guides them toward working through problems themselves. The mentoring approach is grounded in Socratic thinking, where learning happens through prompting deeper questions rather than receiving ready-made responses.

Yukti uses a speech-to-speech architecture with Google Gemini as the underlying model. A layer of context has been added, specific to the course the students are following and the mentorship process the agent needs to support. Students can choose which language to interact in, making it more accessible to a wider group of students. While Yukti can be used in the language students are most comfortable with, the tool also gently encourages the use of English to build confidence ahead of job interviews.

The tool is deployed as a Telegram bot. The team found that students were already active on Telegram and would be unlikely to install a separate app. Students sign up through Telegram, upload their work, such as a project presentation, and receive a callback on their phone as a regular voice call. After the conversation, action items are shared, and students can continue to iterate on their work. Students can access different personas within the bot: one for presentation coaching, another for technical questions, depending on what stage of their project they're at.

Yukti was also used to reach out to students who had not been engaging, to understand what barriers they were facing. This allowed them to collect feedback that would previously have been difficult to gather at scale.

The impact

Yukti is an active pilot currently serving just over 1,100 students. The pilot experienced initial delays due to technical challenges with deploying the voice infrastructure at scale, which coincided with the students' exam period.

The team is collecting data on whether Yukti leads to a measurable difference in project completion rates compared with cohorts without the intervention, and is experimenting to understand the tool's behavioural impact on students' engagement with their learning. As shared by a student from Jaya Engineering College in Tamil Nadu, "I was good at building solutions but never at marketing them. Sometimes we couldn’t even tell whether our project was truly validating the problem statement. Yukti's mentorship was very helpful in this area. As an AI mentor, Yukti made me comfortable in a non-judgmental way, yet always felt like I was talking to a real person. This really boosted my confidence."

Lessons learned

• Meet students where they already are. Deploying through Telegram rather than a standalone app removed a significant friction point. Using a regular phone callback rather than an in-app call made the interaction more accessible. Small decisions about delivery channels can make a meaningful difference to whether students engage.
• Evidence comes before government investment. The state government funds course delivery but does not fund the AI tool. The team is running the pilot at their own cost to generate evidence that the intervention improves outcomes, with the aim of integrating it into the government programme if results support that case.
• AI tools for learning can be designed to support thinking, not replace it. The core design principle of Yukti is that the tool should guide students through a problem rather than solve it for them. This is a deliberate pedagogical choice, and one that distinguishes the approach from general-purpose AI assistants.
• A working prototype does not mean deployment readiness. The team developed a functioning prototype in-house, but deploying it on external infrastructure for real use revealed challenges they had not anticipated. Most voice AI platforms on the market are designed for short, transactional interactions. Using voice AI for longer, more nuanced mentoring conversations where the model needs to reason and respond thoughtfully remains technically limited.

The tool is compliant with India's Digital Personal Data Protection Act (DPDP). Voice data from conversations is handled through tools and architectures that meet the required standards for recording and storage.