Inspiration
ShizenAI was inspired by a problem I have lived through repeatedly: enterprise onboarding feels like being thrown into the deep end with no lifeboat.
At multiple companies, I was handed a maze of wikis, docs, and tribal knowledge and expected to become productive fast. That creates anxiety for everyone involved: new hires are afraid of looking unprepared, managers are afraid to trust them with important work, and HR is left hoping the onboarding process actually works. In practice, onboarding becomes a guessing game.
We wanted to build something that replaces guesswork with confidence. ShizenAI is designed to turn passive documentation into an active learning loop, so organizations can verify understanding instead of assuming it.
What it does
ShizenAI is a role-based training and competency verification platform for enterprise teams.
Admins upload internal knowledge sources, and ShizenAI turns them into structured learning material such as summaries, topics, and flashcards. Employees then go through an active review loop where they answer questions, get evaluated, and are scheduled for future review using spaced repetition.
Instead of asking, “Did they read the docs?”, teams can ask, “Can they actually explain the system back?”
Our goal is simple: give managers peace of mind, give employees a clearer path to mastery, and make onboarding more dependable across the organization.
How we built it
We built ShizenAI as a full-stack web platform with:
- a React + TypeScript + Vite frontend
- a FastAPI backend
- PostgreSQL + pgvector for knowledge storage and retrieval
- Perplexity API for summarization, judging answers, and topic extraction
- optional ElevenLabs support for voice features
- Terraform + GCP for deployment infrastructure
The core flow works like this:
- An admin uploads a knowledge source.
- The backend parses and chunks the content.
- ShizenAI summarizes the content, extracts topics, and generates flashcards.
- Knowledge chunks are stored with vector embeddings in PostgreSQL using pgvector.
- Employees answer review questions.
- The system evaluates answers using retrieval + LLM judging.
- An SRS loop updates review timing so important material is reinforced over time.
We originally explored a more fully local LLM path for tighter inference control, but during the hackathon we pivoted to a Perplexity-first architecture so we could focus on reliability and the product experience instead of wrestling with unstable model hosting.
Challenges we ran into
The biggest challenge was balancing ambition with what could reliably work during a hackathon.
We initially wanted a stronger local-LLM setup running in our cloud environment, but the hosted inference path was not behaving consistently enough for a good demo. Rather than force a brittle architecture, we pivoted toward a more dependable API-first approach and focused on the core value: verifying employee understanding.
Other challenges included:
- designing a clean knowledge-ingestion pipeline
- keeping embeddings and retrieval lightweight enough for hackathon speed
- deciding how strict answer evaluation should be
- building an SRS loop that feels useful instead of punitive
- connecting the technical system back to the very human problem of onboarding anxiety
What we learned
We learned that the strongest products are not just technically impressive; they solve a painful human problem clearly.
On the technical side, we learned a lot about:
- retrieval workflows with pgvector
- FastAPI service design
- shaping LLM prompts for evaluation and summarization
- using Terraform to create a reproducible deployment path
- making pragmatic architecture decisions under time pressure
On the product side, we learned that onboarding is not only a documentation problem. It is a confidence, trust, and verification problem.
What's next for ShizenAI
Next, we want to expand ShizenAI into a more complete voice-first teaching assistant with stronger manager dashboards, better competency analytics, and deeper enterprise integrations.
The long-term vision is to make onboarding feel less like drowning and more like guided mastery.
Built With
- compute
- docker
- elevenlabs
- engine
- fastapi
- gcp
- perplexity
- pgvector
- postgresql
- pydantic
- python
- react
- sql
- sqlalchemy
- terraform
- typescript
- vite
Log in or sign up for Devpost to join the conversation.