VRIKSHA AI - Personal Skill Gardner

Inspiration

The inspiration for this project comes from something we as students face every day. It is easy to get lost in long chat sessions while studying concepts with an LLM, or solve endless coding problems without knowing how much we actually understand. We also miss the chance to talk through our thought process, which is essential during interviews. We wanted to connect learning, speaking, practice and progress in one place, where each session helps us understand what to work on next.

What it does

Our web application helps users study concepts, prepare for interviews and stay motivated towards their goals. Six different AI personas guide users through learning, technical interviews, behavioral interviews and accountability. The platform remembers previous sessions and uses assessed answers to identify areas that need improvement.

This progress is shown through a living 3D Skill Garden, where each plant represents a concept and grows based on what the user demonstrates. Users can inspect the evidence behind their progress, see which concepts need more practice and continue with a coach that has context from their earlier sessions.

How we built it

To build this we designed a "Skill Garden" which is our knowledge engine using 6 different ElevenLabs voice agents with different skills and tools. This knowledge engine stores and grows information and past context related the users studies and interview preparations as a dynamic knowledge base which gets updated after each session of study or preperation. This web app is a voice first platform where real time voice agents are configured and deployed (ElevenLabs). Custom webhooks were configured to set up dynamic offline workflows on audio and text. We used tiger data as our database engine to power real time analytics and garden updates.

For processing uploaded study materials we built a document ingestion pipeline which extracts text, divides it into smaller searchable chunks and maps them to concepts in our curriculum. These chunks are stored with their source references so our voice agents can retrieve relevant information and ask questions based on the users material.

We used Hatchet for background workflows such as document processing, post-session assessment and scheduled accountability check-ins. This lets us process completed sessions separately from the live conversation without making the user wait for the entire assessment.

For evaluating the users answers we used a separate LLM judge. We also track whether the user answered independently or needed hints, so assisted answers and independent answers are treated differently.

We built a rules engine which takes these assessments and decides how the garden should grow.Each growth event is linked to the evidence behind it, and repeated processing of the same session does not give the user additional growth.

Our memory persistence engine stores past assessments, misconceptions and concepts which need further practice. Before a new session this information is used to prepare a relevant briefing for the voice agent, helping it continue from the users previous session and focus on areas where they need improvement.

For technical interviews we built a code editing workflow with sandbox execution using E2B. This lets us assess the users explanation along with their code and test results. For behavioral interviews we used structured questions and follow-ups to assess how the user explains their experiences.

We also built an owner dashboard to track users, sessions, usage, costs and assessment failures. An AI analytics assistant uses a restricted MCP gateway to query approved data from Tiger Data and ElevenLabs. Authentication and permission checks control access, while usage tracking and budget reservations help manage spending across the different services.

Challenges we ran into

We ran into 1) Designing our database schema to support pgvector retrieval alongside user memories and assessments. 2) Keeping Skill Garden fast during live conversations while processing detailed assessments in the background. 3) Making garden growth reflect actual understanding, distinguishing independent answers from hinted answers and preventing duplicate growth.

Accomplishments that we're proud of

1) Keeping voice agents responsive while handling complex backend workflows through careful system design. 2) Building a visually motivating, interactive 3D garden that shows progress and connections between concepts. 3) Creating a live coding interview experience where users can explain their approach, write code and test their solutions.

What we learned

1) How to connect real time voice agents with persistent memory and background assessment. 2) Why tracking evidence and assistance matters when measuring learning progress. 3) How to turn complex learning data into an interactive visual experience.

What's next for MHacks

1) Test with students and improve coaching based on their feedback. 2) Expand beyond computer science into more subjects and interview roles. 3) Refine assessments and follow-up practice to better track long-term retention.