AI · SaaS · Dashboard · 2026
DeshiGPT
A portfolio-grade multi-model AI chat workspace: 8 streaming models, OCR studio with confidence scoring, web-grounded answers, projects, artifacts, voice in/out, and a usage dashboard.
- Models
- 8
- Streaming latency
- <200ms
- Workflows
- Chat · OCR · Web · Voice
- Client
- Self-initiated demo
- Year
- 2026
- Timeline
- 4 weeks
- Categories
- AISaaSDashboard
- Role
- Full-stack engineer & product designer
- Services
- Product designFull-stack developmentAI integrationPrompt engineering
- Stack
- TanStack StartReact 19TypeScriptTailwind CSS v4Vercel AI SDKStreamdown
The challenge
Most ChatGPT clones lock you into one model and skip the workflows that make AI actually useful — document OCR, grounded web search, project memory, and voice. The goal was to ship a single workspace that combines them with production-grade streaming UX.
My approach
Built on TanStack Start with server functions handling auth, streaming, and tool calls. Routed 8 frontier models through a unified AI gateway, layered OCR with confidence scoring, and added web-grounded answers, projects, artifacts, and voice in/out — all with a real-time usage dashboard.
Key features
- 8 streaming AI models in one workspace (GPT, Claude, Gemini, Llama, and more)
- OCR studio with per-token confidence scoring
- Web-grounded answers with citations
- Projects and artifacts for organized, reusable work
- Voice input and text-to-speech replies
- Usage dashboard with per-model token tracking
Outcome
A polished, fast, multi-model AI workspace that demonstrates end-to-end product thinking: streaming chat, document intelligence, grounded answers, and voice — wrapped in a clean dashboard suitable for real users.