SmartDuke builds rag systems for education and EdTech — systems that ground AI answers in your own data, with citations and freshness guarantees. AI tutors fundamentally change the economics of personalized education. Per-student tutoring cost drops by orders of magnitude — but only if the AI is actually grounded and pedagogically sound. Our approach: Retrieval evals before any LLM call, citation-rate monitored as a first-class metric, and explicit refusal when confidence falls below threshold.
The problems
we keep solving.
AI tutors fundamentally change the economics of personalized education. Per-student tutoring cost drops by orders of magnitude — but only if the AI is actually grounded and pedagogically sound.
Personalized learning at scale is expensive
Per-student tutoring is the gold standard but cost-prohibitive at any meaningful scale.
Assessment generation is repetitive busy-work
Creating quizzes, exams, and exercises that align with curriculum is high-effort and low-leverage.
Curriculum adaptation requires constant manual work
Adapting content to grade level, language, and learning style is rarely scoped into product timelines.
Three things we'd build
first.
Concrete starting points for rag systems in education and EdTech. We pick the one with the highest leverage and the cleanest measurement story.
- Idea 01
Personalized tutor agent that adapts to student level, surfaces hints, and tracks misconceptions over time
- Idea 02
Assessment-generation copilot that produces quiz questions calibrated to curriculum and difficulty
- Idea 03
Content-adaptation system that translates and adjusts existing materials for new audiences
Production-grade,
from day one.
Hybrid retrieval (BM25 + semantic + reranker), source-priority ranking, freshness gating, and answer-capsule generation with sentence-level citations.
Evals before launch.
Every loop, tool call, and structured output is graded with a frozen test set and an explicit rubric. Failed evals block the deploy.
Telemetry from day one.
Traces, latency budgets, token costs, and error rates wired up before the first user touches the system.
Guardrails as architecture.
Input validation, output verification, escape hatches, and human handoff paths designed in — not bolted on after incidents.
Boring stack on the edges.
Cutting-edge model in the middle. Reliable infrastructure around it. Stability where it earns its place.
- ×Hallucination on long-tail queries
- ×Stale documents silently used
- ×Missing or wrong citations
- ×Retrieval that fails on the queries that matter most
Common questions.
01How long does it take to build rag systems for Education?
How long does it take to build rag systems for Education?
Discovery is one week. A working prototype (Spark) is 2–3 weeks. Full production Build for education and EdTech typically runs 8–12 weeks depending on data complexity, integrations, and compliance scope. We commit to a precise timeline at quote stage.
02What does pricing for rag systems typically look like?
What does pricing for rag systems typically look like?
Every engagement is scoped to outcomes, not hours. Discovery starts in the low four figures. Spark and Build are priced per project. Embed retainers are monthly. We return a quote within 24 hours of inquiry.
03Can you take over an existing RAG project that's stalled?
Can you take over an existing RAG project that's stalled?
Yes — it's a common engagement. We review what's there, tell you honestly what stays and what we'd rebuild, then ship it to production. Education engagements often start this way.
04What's different about your approach to rag systems?
What's different about your approach to rag systems?
Retrieval evals before any LLM call, citation-rate monitored as a first-class metric, and explicit refusal when confidence falls below threshold. We hold the same engineering bar across every engagement, regardless of industry — but the specifics for education and EdTech are tuned to your trends and pain points.
Same capability, different industry.
Same industry, different capability.
Have an AI product
that needs to ship?
Tell us where you are — early concept, broken prototype, or scaling something that already works. We'll come back within 24 hours with a take and a quote.