The AI attendance agent that replaced manual roll calls Schools still take attendance the same wayThe AI attendance agent that replaced manual roll calls Schools still take attendance the same way
The network for creativity
Join 1.25M professional creatives like you
Connect with clients, get discovered, and run your business 100% commission-free
Creatives on Contra have earned over $150M and we are just getting started
Back to feedPost
Taimoor's avatar
Taimoor Khan
pro
• Jun 30
The AI attendance agent that replaced manual roll calls
Schools still take attendance the same way they did 30 years ago. A teacher calls names. A student says "present." Someone marks a register. Multiply that by every class, every period, every day, and you've got hours of dead time baked into the system.
When I built the AI Attendance Agent for EduPilotPro, the goal was simple: make attendance autonomous. No teacher input. No manual marking. Just an agent that handles it.
Here's the part that made the architecture interesting.
Why two LLMs?
The agent runs on Google Gemini as the primary model. Gemini handles the core reasoning: interpreting attendance data, resolving edge cases (late arrivals, partial attendance, substitutions), and generating structured outputs that feed directly into the school's system.
But relying on a single cloud LLM for a school environment is a risk. API latency spikes. Rate limits during peak hours when every classroom hits the system at once. Outages that turn your "autonomous" agent into a blank screen.
So I added WebLLM as a fallback layer. WebLLM runs inference directly in the browser. If Gemini is slow or unreachable, the agent degrades gracefully to local inference instead of failing entirely.
The trade-off is real:
Gemini gives you stronger reasoning and richer context windows, but you're dependent on network + API availability
WebLLM gives you resilience and zero-latency local inference, but with a smaller model and less sophisticated outputs
For attendance (which is structured, repetitive, and doesn't need GPT-4-level reasoning), the fallback model handles it fine. The user never notices the switch.
The stack underneath
Node.js on the backend. PostgreSQL for persistence. The agent writes attendance records directly to the database with full audit trails, so administrators can trace every decision the AI made and why.
No fancy orchestration framework. No LangChain. Just clean Node services that call the right model, validate the output, and write to Postgres.
What I'd tell founders building AI agents
Your agent will go down. Plan for it. The "two LLM" pattern (cloud-primary, local-fallback) works well for any task where the fallback model is competent enough to handle the 80% case. Attendance qualifies. Complex medical reasoning probably doesn't.
Build the fallback before you need it. You won't have time when Gemini returns a 429 at 8:45 AM on a Monday.
→ Full case study: EduPilotPro AI Attendance Agent
AI Agent Engineerllmlargelanguagemodel
Back to feed
The network for creativity
Join 1.25M professional creatives like you
Connect with clients, get discovered, and run your business 100% commission-free
Creatives on Contra have earned over $150M and we are just getting started