Generative AI

Multimodal AI Chat Platform

A chat platform handling text, image, voice and document inputs across three frontier models, built in eight weeks.

Next.jsTypeScriptOpenAIAnthropicLangChain

At a glance

What this shipped

The numbers that mattered to the client, measured before and after.

<800ms
First-token latency
−38%
Cost / conversation
8 weeks
Build duration
4
Models supported

The problem

What we were called in to fix

The client, an AI startup, needed a single product where users could chat across modalities (paste an image and ask about it, transcribe a voice note, upload a PDF and query it) using whichever of GPT-4o, Claude or Gemini was best for the job.

Their existing prototype could do one model with text-only. Latency was several seconds, conversation memory was broken, and adding a new modality meant rewriting half the stack.

Our approach

How we actually built it

No magic. Just the right architectural calls in the right order.

We rebuilt the message pipeline around a provider-agnostic interface so model selection, model swap and cost routing all happen in one place.

Streaming-first architecture: tokens render the moment the LLM emits them. No spinner-on-spinner experience.

Modality handlers (text, image, audio transcription, document RAG) compose cleanly: each one is a small, testable module rather than a special case in the chat loop.

Conversation memory is durable and queryable in Postgres, with a summary-on-overflow strategy so long threads stay coherent without blowing the context window.

The outcome

What changed for the client

First-token latency under 800ms across all three providers, end-to-end.

Added Claude 3.5 Sonnet as a fourth model behind the same interface in under two days.

Reduced inference cost per conversation by 38% via smart model routing: cheap models for cheap turns, frontier models when complexity demands it.

Tech stack

Every meaningful piece

Next.js 15TypeScripttRPCOpenAIAnthropicGoogle GeminiLangChainpgvectorPostgreSQLRedisAWS

“We don't do generic case-study writeups. Want the unredacted version with names, screenshots and architecture diagrams? We share those on a call.”

Yours could be next

Have a project that needs shipping?

Send us a short brief and you'll have a clear scope, fixed quote and timeline within 24 hours.