Preview

GT-ASSIST-V2

One Kenyan account for multi-model AI chat, image, video and voice — billed in KES.

Live demo

The problem

A Kenyan team that wants AI in its daily work ends up juggling half a dozen separate tools — one for chat, one for images, one for video, one for transcription — each with its own USD subscription, its own login, and no way to pay in shillings. Usage is impossible to budget, and most tools cannot even be trialled without a foreign card. GT-ASSIST-V2 puts the whole set behind a single account with one credit balance and local payment rails.

What the agent does

  1. 1The user signs in and receives a daily free credit allowance; top-ups go through a Paystack checkout they confirm themselves.
  2. 2They send a chat message; the app routes the request to a model suited to the job.
  3. 3The reply streams back token by token, with tool calls (web search, knowledge base) surfaced inline.
  4. 4Media requests branch off the same credit balance: images, image edits, video animation, speech-to-text and text-to-speech.
  5. 5Conversations, media and the credit ledger sync to the user's account so any device resumes where they left off.
  6. 6Rate-limited provider calls retry with exponential backoff and jitter rather than failing the user's request.

Agent architecture

Models

  • DeepSeek V3.2 / V4-Flash (fast everyday chat, 1 credit)
  • Kimi K2.6 / K3 (long-context work)
  • GLM-5.1 / GLM-5.2 and Qwen3 235B-Thinking (hard reasoning lane)
  • Qwen-Image-2512 + FLUX.1 schnell (images), LTX 2.5 (i2v/t2v video), Whisper (STT), Kokoro TTS

Tools

  • Chutes API (chat, image, video, audio)
  • Clerk (identity) and Supabase (cloud sync, edge functions)
  • Paystack (KES payments, webhook confirmation)
  • Web search + knowledge-base grounding (searchAgent / knowledgeBase)
  • retryFetch (429/503 retry with exponential backoff + jitter)

Memory

Supabase persists conversations, generated media and the per-user credit ledger across devices; the conversation branch/edit history is part of that state.

Routing

The user picks a model per task, and the platform groups them by cost and depth: 1-credit lanes (DeepSeek V4-Flash, Qwen3.6-27B) for everyday chat, reasoning lanes (Kimi K3, GLM-5.2, Qwen3-235B-Thinking) for hard work, with media models routed by task type.

Hosting

Provider: Vercel

Plan: Hobby

Region: iad1

URL: https://create.gashotech.com

Tier 1 checklist

5 of 12 passing

Every project must clear all 12 Tier 1 checks before it is presented as live. This is enforced in code, not by convention.

  • README explains the architecture in ≤ 5 minutesMissing — README has no architecture section yet.
  • 60-second demo video recordedMissing — 60-second walkthrough not recorded.
  • Public GitHub repositoryMissing — repo is private pending a user call and a secret sweep.
  • Prompts versioned
  • Every model call loggedUnverified — no per-call log surfaced yet.
  • Tool outputs schema-validatedUnverified — tool output schemas not demonstrated.
  • Retries with backoff
  • Secrets / PII stripped before the modelFAILING — the live bundle ships a Chutes API key inlined by Vite. Fix: rotate the key and proxy Chutes calls server-side.
  • Human approval on money, email and delete actions
  • Streaming responses
  • Repeat queries cachedUnverified — no repeat-query cache demonstrated.
  • Token / time / cost budget cap

Not presented as live yet — 7 of 12 Tier 1 checks outstanding.

All ProjectsBack to Home