PromptPilot 🚀

Bridging the Gap Between Raw Intent and Production-Ready Prompts.

PromptPilot is an advanced, research-grounded Prompt Engineering Agent designed for non-technical professionals. It eliminates the "trial-and-error" loop of working with LLMs by using an agentic reasoning workflow to transform vague thoughts into structured, high-performance instructions.

---

🎯 The Problem

Most professional users struggle with Instruction Drift and Prompt Ambiguity. While frontier models are powerful, they require specific structural markers (XML, CoT, Delimiters) to perform consistently. PromptPilot acts as the "Navigator," translating simple language into the technical "handshake" LLMs require.

✨ Key Product Features

• Agentic Interviewer: Uses a "Gap Analysis" logic to identify missing variables (Context, Persona, Format) and asks targeted follow-up questions before generating.
• Knowledge Vault (RAG): A curated library of 2026 prompt engineering research. Every prompt is grounded in techniques like Chain-of-Thought, Chain-of-Density, and Self-Consistency.
• Asymmetric Reasoning: Powered by Gemma 3n E4B, utilizing Matryoshka embeddings and Per-Layer Embedding (PLE) caching for high-density logic with sub-400ms latency.
• Power Mode: Provides a transparent "Reasoning Trace" (<thinking> tags), showing the user exactly how the AI interpreted their request.
• Model-Aware Optimization: Tailors output structure specifically for the target model (Claude, GPT-4, Gemini, or Grok).

---

🏗️ Technical Architecture

The Intelligence Stack

• Core Logic: google/gemma-3n-e4b-it (optimized for latency-to-logic efficiency).
• Vector Database: Supabase (pgvector) storing 1024-dimension embeddings.
• Embedding Model: intfloat/multilingual-e5-large-instruct (utilizing passage:/query: instruction prefixes).
• Semantic Cache: Upstash Redis to reduce COGS and latency for redundant high-intent queries.

Evaluator-Optimizer Design Pattern

PromptPilot doesn't just "guess." It follows a closed-loop system:

1. Ingestion: RAG retrieval of best practices.
2. Synthesis: Gemma 3 generates the "Improved Prompt."
3. Audit: Internal regression testing against a G-Eval rubric (Faithfulness, Specificity, Structure).

---

📈 Performance & Unit Economics

• Latency: Average Time to First Token (TTFT) < 350ms via Together AI Serverless.
• Context Grounding: 100% of generated prompts include citations from the Knowledge Vault.
• Optimization Alpha: Average 25-35% reduction in "Prompt Drift" compared to raw user inputs.

---

🛠️ Getting Started

Prerequisites

• Node.js 20.6.0+
• Together AI API Key
• Supabase Project (with pgvector enabled)
• Upstash Redis (for semantic caching)

Installation

1. Clone the Repo:

   git clone https://github.com/SriramGanne/promptpilot.git
   cd promptpilot

2. Environment Setup: Create a .env.local file:

   TOGETHER_API_KEY=your_key
   NEXT_PUBLIC_SUPABASE_URL=your_url
   NEXT_PUBLIC_SUPABASE_ANON_KEY=your_key
   UPSTASH_REDIS_REST_URL=your_url
   UPSTASH_REDIS_REST_TOKEN=your_token

3. Seed the Vault:

   node --env-file=.env.local scripts/ingest_research.mjs

4. Launch:

   npm run dev

---

🗺️ Roadmap

• [ ] Multimodal Intent: Support for image-to-prompt (Visual Prompt Engineering).
• [ ] Team Workspaces: Collaborative Knowledge Vaults for enterprise teams.
• [ ] Live Eval Dashboard: Public-facing metrics on prompt "Win Rates" using Sonnet 4.6 auditing.

📄 License

Distributed under the MIT License. See LICENSE for more information.

---

Developed by Sriram Ganne Senior AI Product Management Portfolio Project