NexaGear Shield: Enterprise AI Framework

A deterministic governance proxy and semantic firewall that secures enterprise Generative AI from policy hallucinations, fraud, and prompt injections.

Comment

Inspiration 💡

Generative AI is revolutionizing conversational interfaces, but enterprises are terrified to deploy it. Without strict guardrails, standard LLMs hallucinate policies, offer unauthorized discounts, and violate core business rules—exposing companies to significant financial and legal liabilities. We were inspired by the urgent industry need to transform "non-deterministic" chatbots into "deterministic," goal-aligned agents that managers can actually trust.

What it does ⚙️

NexaGear Shield is an Enterprise AI Controller and Governance Middleware. It acts as an absolute authority proxy between the user and a Large Language Model.

  1. Live Business Engine: Managers can update business constraints (e.g., maximum discount caps, return policies) in real-time via a secure Ops Dashboard. The constraint state is instantaneously synced to the AI.
  2. Explainability Ledger: Every AI response is audited. The system generates transparent, timestamped "Explainability Tokens" that visually prove exactly which policy triggered a restricted AI response.
  3. Semantic Firewall: We implemented a cybersecurity layer that scans the user's intent vector before LLM execution. If a Prompt Injection or Jailbreak attempt is detected, the API request is killed instantly and a severe fraud alert is logged.

How we built it 🏗️

We engineered a high-speed B2B architecture:

  • The Brain: We integrated the Google Gemini 2.5 Flash model, heavily tuned with low temperatures to enforce logic compliance.
  • The Interceptor: We built a Python FastAPI backend to act as the in-memory proxy. This layer handles the Live State Vector (Business Rules), intercepts prompts, validates outputs, and extracts the generated audit tags.
  • The Command Center: The frontend is a highly responsive HTML5/Vanilla JS Single Page Application (SPA). It features a split-view interface showcasing the Consumer Endpoint alongside the Manager's Live State Controller, styled with modern Glassmorphism.

Challenges we ran into 🛑

Our primary challenge was instructing the LLM to understand logical boundaries without breaking character. Initially, the AI would either completely forget the business rule or aggressively break the conversation flow. We overcame this by engineering a robust systemic injection architecture that maps the live variables directly into the hidden system prompt prior to generation. Additionally, extracting the exact trigger tokens cleanly required us to build a custom string-validation layer in our Python middleware.

Accomplishments that we're proud of 🏆

  • Successfully achieving Zero-Latency State Updates: A manager can click "Block Returns" and the next customer message is instantly restricted without any server restart or model fine-tuning.
  • Building the Semantic Firewall: We are incredibly proud of our cybersecurity layer that actively defends against malicious prompt engineering and saves compute costs.
  • Developing an Enterprise-Grade UI: Our animated SPA transition and telemetry data visualizations make the product feel like a polished, million-dollar enterprise platform.

What we learned 🧠

We learned that GenAI alone is not a product; it is an engine. The real value for enterprises lies in the **governance architecture holding

Built With

Share this project:

Updates