Apurv Mehra | Portfolio

The premise

India produces 1.5 million engineering graduates a year. Seven in ten are deemed unemployable. The standard response is always more skilling: better content, sharper assessments, faster certifications.

After piloting programs with 5,000+ students in Tier-2/3 colleges and 1,000+ at IITs, IIITs, and BITS, I'm convinced the upstream bottleneck is different. It's not a deficit of skill. It's a deficit of agency: the capacity to see one's options, update beliefs, and make deliberate choices about one's own future.

After five years of running into the same wall from different angles, I now frame it as a chain: Information → Assimilation → Trust → Agency → Action. Most skilling platforms skip straight to action. The chain breaks most often at trust.

Three projects taught me this directly. Two adjacent ones reinforced it.

The evidence

1. Know Thy Choice: You can't choose what you've never seen

Know Thy Choice built experienceships: short, project-based career exposures under industry mentors for high-school students.

Students cannot aspire to careers they've never seen.
Diagnostics produce labels, not insight. Being told "you'd suit UX" changed nothing. Doing UX work for a week changed everything.
Iterative counselling starting in Grade 9 produced fundamentally different decisions than default-driven choices in Grade 12.

This is where the chain first became visible. Information alone (a psychometric report, a career list) doesn't create agency. It has to be assimilated through lived exposure before a student can act on it.

2. Sakshm AI: Scaffolding that transfers, not replaces

Sakshm AI, with BITS Pilani, IIIT Delhi, and Microsoft Research India, deployed "Disha," an AI tutor that refuses direct answers and guides through Socratic questioning. 3,951 registered users, 1,170 active. (Paper | Press)

Three findings:

Self-weaning. Highly engaged users (Q3) leaned on AI 30.8% of the time. Super-users (Q4) only 12.7%. As proficiency grew, AI reliance dropped. This is the opposite of the dependency narrative. When designed well, AI scaffolding transfers capability rather than replacing it.

The Socratic ceiling. Students embraced guided questioning for easy/medium problems but bailed to ChatGPT on hard ones. Scaffolding only works when the learner has enough foundation to reason from. Without that foundation, the assimilation step hasn't happened yet.

Cultural trust. "Disha" (Hindi for "direction"), a deliberately Indian persona, reduced friction in ways a generic Western-named bot could not. Cultural design is not a footnote; it's how trust gets built.

We also open-sourced a Bayesian A/B testing framework from this deployment: 1,186 participants, 6.35% engagement lift during the "discussion phase" where students assimilate and articulate before solving.

3. Aspireworks: Beliefs shift before outcomes can

Aspireworks, our AI career-navigation platform, deployed across two Karnataka government colleges (200+ students) in partnership with the Dept. of Higher Education and Microsoft Research India, plus 1,000+ students through a multi-state Codathon.

The govt-job default is a belief, not a choice. 29% of students named civil services as their aspiration. Few could articulate why. Most reported it as the default expectation in their families. After exposure to alternative pathways, preferences shifted measurably.

Channel > content. What drove adoption wasn't platform quality. It was the college's endorsement. Students engaged because faculty framed it as part of their education. Government-channel distribution is a trust strategy, not just a scale strategy.

This is the clearest demonstration of the chain in action: institutional trust (step 3) unlocked engagement, which enabled belief-shifting (step 2), which expanded choice (step 1), which produced agency (step 4).

What adjacent work reinforced

Two other projects shaped this thinking, even though they weren't directly about student agency:

BlendNet (Bihar). Our COMPASS 2024 paper documented 258 retailers reaching 68,000+ end-users in three months as digital service intermediaries (Project). The finding: the barrier to digital adoption wasn't access but confidence that the service would help. Local retailers delivered trust, not the service itself. This is why every product we've built since includes a human in the loop.

Comuniqa. With Microsoft Research India and IIIT Delhi, we compared LLM-only, human-expert-only, and combined coaching for English speaking skills (Paper | Project). The LLM was accurate but lacked empathy. Learners felt less understood. The combined group produced the strongest outcomes. The lesson that stuck: AI mass-customizes, humans mass-encourage. You need both.

The framework

After five years, this is how I think about it:

| Step | What it means | Where it breaks | |------|--------------|-----------------| | Information | Encounter options you hadn't considered | Choice space stays artificially small | | Assimilation | Make sense of it, update beliefs | Most platforms skip this entirely | | Trust | Confidence that acting will help | Without it, even understood information stalls | | Agency | Capacity for deliberate choice | Cannot exist without steps 1-3 | | Action | Build capability, convert to outcomes | Where most "skilling" starts and fails |

Most Indian edtech jumps to step 5. The chain breaks at step 3.

What remains unproven

We have evidence that:

AI scaffolding transfers capability without creating dependency (Sakshm AI)
Career exposure shifts what students believe is possible (Know Thy Choice)
Institutional trust unlocks engagement that content quality alone cannot (Aspireworks)
Trusted human intermediaries are foundational to adoption in low-resource contexts (BlendNet)
AI + human combinations outperform either alone for outcomes (Comuniqa)

We still need to demonstrate:

Whether shifted beliefs translate into shifted outcomes (placement rates, salaries, time-to-employment)
Whether the chain can be delivered end-to-end at state scale
Whether outcomes differ meaningfully for women, first-generation graduates, and Tier-3 students

References

Sakshm AI: Disha: A Socratic AI Tutor for Programming Education. ACM COMPASS 2025. Paper | Project
Comuniqa: Exploring Large Language Models for Improving English Speaking Skills. ACM COMPASS 2024. Paper | Project
BlendNet Last-Mile Service Delivery: ACM COMPASS 2024. Paper | Project
Sakshm AI LLM Evaluation Framework: Bayesian A/B testing middleware, MIT-licensed. GitHub
Know Thy Choice: Career exploration platform for school students in India. Website