Software Engineer

EduAgent Platform

AI tutoring agents for education

2026 — Present·6 engineers

0k+

learner sessions / day

questions auto-resolved

<0.0s

p95 first-token latency

−0%

token cost via caching

Overview

An AI agent platform for education at SiliconPrime: LLM-powered tutoring agents answer learner questions over course materials using RAG, orchestrate tools (quiz generation, progress lookup, curriculum search), and stream responses in real time to web and mobile clients.

Challenges

Answering course-specific questions accurately without hallucinating beyond the curriculum.
Keeping streaming latency low while agent runs involve multiple tool calls and retrieval rounds.
Controlling LLM token spend as daily sessions grew past 5k.

Solutions

Built a RAG pipeline over versioned course materials with pgvector embeddings and metadata filtering per course and grade level.
Designed an agent orchestration service with parallel tool execution and an SSE streaming gateway that flushes first tokens before slow tools finish.
Added Redis-backed prompt/response caching and context-window pruning keyed on curriculum version.

Results

85% of learner questions resolved without human tutors, measured over 5k+ daily sessions.
p95 first-token latency held under 1.2 seconds during peak evening study hours.
Token cost per session cut ~35% after caching and pruning rollout.

Architecture

Tech stack

Node.js
NestJS
PostgreSQL
pgvector
Redis
LLM APIs
SSE
WebSocket
Docker
AWS

Process

01
Agent & RAG design
3 weeks
Modeled agent tools, retrieval flows, and curriculum embedding strategy.
02
Platform build
Ongoing
Implemented orchestration services, streaming gateway, and evaluation harness.
03
Cost & latency tuning
Continuous
Caching, pruning, and load testing against session growth.

Next project

Electronic Medical Record

Multi-tenant clinic & hospital SaaS