Prompt Asset & Evaluation Platform

v1.0 · Production

PromptDeck

PromptDeck is a prompt asset and evaluation platform for AI teams. Today, most teams' prompts are scattered across code constants, note apps and chat history: a one-word change ships with zero regression testing, quality is judged by gut feeling, and every release is a gamble. PromptDeck upgrades prompts into version-controlled engineering assets — every change gets a git-style diff, every version runs against your own case sets, and score regressions block merges right in CI. Compare local Ollama models against the cloud Claude API side by side, with the whole stack available for self-hosted deployment.

Open Consolev1.0 · Production

Problems We Solve

  • Prompts scattered across code, notes and chats — no single source of truth
  • No regression tests for changes; 'feels better' goes straight to production
  • Model switches and upgrades cause unquantifiable quality regressions
  • Sensitive eval data cannot be uploaded to third-party SaaS

Core Features

Prompt Version Control

Git-style history and word-level diffs. Every prompt is a first-class asset with branching and rollback.

Dataset-driven Evaluation

Build your own case sets (input + expected behavior); an LLM judge scores automatically with fully traceable reports.

Multi-model Comparison

Same prompt, same dataset — local Ollama (qwen / deepseek-r1) and the cloud Claude API benchmarked side by side.

Regression Gate

CI integration: PRs trigger evaluations, and scores below baseline block the merge. Prompt changes finally get a safety net.

Evaluation Worker Queue

Evaluations run asynchronously in a queue — large case sets never block the console, with observable progress and retries.

Self-hosted Deployment

One Docker Compose command. All data lives in your own PostgreSQL; inference can run entirely on in-network GPUs.

System Architecture

Request path, top to bottom

Console
Next.js

Prompt editing, diff views, evaluation reports and comparison dashboards

API
FastAPI

Version management, evaluation orchestration, CI webhook gates

Inference
Ollama (local RTX 5090) + Claude API

Private evaluation on local models, cloud models as reference baseline

Data
PostgreSQL + worker queue

Persistent prompt versions, case sets and eval records; async evaluation jobs

Tech Stack

Next.js 15 consoleFastAPI backendPostgreSQL storageOllama local inference (qwen / deepseek-r1)Claude API cloud baselineEvaluation worker queue

Pricing Direction

Early Access
Free

Full features for invited teams, in exchange for real-world evaluation feedback

Team Subscription
Per-seat (planned)

Hosted version billed per team seat, evaluation quota included

Self-hosted License
Annual license (planned)

Deployed inside your network, integrated with your own GPUs and models

Your prompts deserve regression tests

Early access seats are limited — priority goes to teams with LLM features already in production.

Apply for Early Access