The challenge
Yaapt, an academic social platform where students share their college experiences, holds a large, fast-growing multi-modal content library. Its AI features called external LLM APIs directly, the infrastructure was self-managed, and search quality couldn't keep up with what students published. Yaapt needed a foundation that could deliver sub-second AI response, govern AI cost and behavior at a single surface, and search across both the visual content and the text inside it.
What we built
All AI orchestration moved to AWS Bedrock, one governed surface for every model call, with cost, latency, guardrails, and model choice managed in one place. Multi-modal search was rebuilt so every image carries complementary embedding representations, capturing both fine-grained visual structure and high-level semantics. Text inside images is extracted and indexed through the same pipeline as the visual content, so a single query covers both. Yaapt's database, cache, and vector search moved off self-managed infrastructure onto AWS managed services, and the platform runs on a production-ready architecture of asynchronous microservices designed for scale-out.
Why this matters
Students get sub-second, multi-modal search that understands both the imagery and the text within it, while every AI feature runs through one governed Bedrock surface with consistent observability and cost discipline. Moving off self-managed infrastructure onto AWS managed services cut operational overhead and left headroom to scale. Yaapt approved the production migration, with a modernization roadmap, architecture diagrams, runbooks, and Infrastructure-as-Code handed over at closeout.