Meera
An agentic AI front door for university service desks: concerns arrive in plain language, get parsed and classified, are resolved automatically where possible, and routed to the right team when not. Top 3 of 100+ teams at the KPMG Academic Innovation Challenge.
- React
- Next.js
- Cloudflare D1
- Cloudflare R2
- Groq
- Ollama
- Deepgram Aura
Problem
University service desks drown in unstructured requests — emails, forms, walk-ins — each phrased differently and routed by hand. Simple, answerable questions wait in the same queue as genuinely hard cases, so response times suffer and staff burn time triaging instead of resolving.
Approach
We built Meera as an agentic intake layer. A request arrives in natural language (typed or spoken), and the system parses intent, classifies the concern, attempts an automated resolution from known answers, and only escalates to the appropriate human team when it can't resolve the case itself. Voice intake is transcribed in real time so the same pipeline serves phone and walk-in channels.
Technical Decisions
Agentic classify-then-resolve pipeline
Instead of a single LLM call, requests flow through discrete stages — parse, classify, attempt resolution, route. Separating the stages made each one debuggable under hackathon time pressure and let us swap models per stage (fast Groq inference for classification, local Ollama for cheaper drafting).
Voice intake with Wispr + Deepgram Aura
Service-desk concerns often come in by voice. We layered Wispr and Deepgram Aura for low-latency transcription so spoken requests feed the exact same parse → classify → resolve path as typed ones, rather than building a separate voice flow.
Tradeoffs
Automated resolution vs. safe escalation
We tuned the agent to escalate aggressively when confidence was low. Fewer requests get fully auto-resolved, but the ones that do are trustworthy — the right trade for a support context where a wrong automated answer is worse than a short wait.
Hackathon shipping vs. production hardening
Built and demoed inside the challenge window, so retries, auth, and audit logging were scoped to the demo path. The next iteration would harden the escalation routing and add per-team analytics.
What I learned
- 01
Splitting an agent into explicit stages beats one giant prompt when you need to debug and demo under pressure.
- 02
Matching the model to the stage (fast vs. cheap vs. local) keeps an agentic system affordable.
- 03
In support tooling, knowing when not to answer is as valuable as the automated answer itself.