AI Project Assessor: Enterprise Review Engine

The Problem

Enterprises are struggling to evaluate AI projects consistently before committing funding. Without a centralised, multi-disciplinary review process, organisations risk green-lighting projects that lack strategic alignment, fail to meet compliance and privacy standards (like cross-border data transfer rules), or hide architectural flaws and escalating token costs.

We need a pre-build review layer that standardises AI project assessments across business, technical, financial, and legal dimensions.

Governance and Risk Framework

The AI Project Assessor enforces a structured review process by analysing proposed AI use cases through five distinct professional lenses. This multi-hat approach ensures a holistic evaluation of feasibility, risk, cost, and compliance—forcing stakeholders to confront potential pitfalls before any code is written or data is processed.

Key Governance Pillars

To ensure rigorous oversight and seamless delivery, every project is evaluated against:

1. CEO Lens (Strategic & Business Value): Evaluates risk-vs-value, strategic alignment, and designates required executive sign-offs.
2. AI Solution Architect Lens (Technical Feasibility): Assesses technical viability, token efficiency, architecture patterns, and critical drift risk or failure modes.
3. Product Manager Lens (User Value & Measurement): Ensures user-centricity, defines exact KPIs with thresholds, and evaluates human-in-the-loop requirements.
4. Accountant / Finance-Ops Lens (Cost & ROI): Provides conservative, likely, and aggressive token-usage and cost scenarios, combined with FTE ROI comparisons.
5. Data Privacy Officer Lens (Compliance & Privacy): Flags PII risks, identifies control gaps, and specifically mandates cross-border data transfer review mechanisms (e.g., TIAs, SCCs).

Architecture

The application is built on a modern, full-stack React and Express architecture, optimised for rapid assessment generation.

• Client-Side: A responsive React interface using Tailwind CSS, structured for capturing nuanced project briefs, usage expectations, and compliance constraints. It includes pre-configured presets for common enterprise scenarios.
• Server-Side: An Express.js backend that securely proxies requests to Google's Generative AI models.
• API Integration: Validates inputs and translates them into a highly structured prompt, utilising JSON Schema output constraints to guarantee a consistent, predictable assessment object.

Agent Orchestration

Currently, the system utilises a highly constrained large language model (gemini-2.5-pro) prompted to adopt a multi-persona "mixture of experts" strategy internally.

The agent orchestration follows a strict sequence:

1. Ingest unstructured or semi-structured user requirements.
2. Estimate missing data (e.g., token volume, costs) using explicitly stated assumptions.
3. Execute evaluations across the five governance pillars.
4. Perform a final "Prompt Engineer-style" synthesis to eliminate contradictions and ensure an executive-grade tone.
5. Return guaranteed JSON matching a strict TypeScript interface for UI rendering.

Tech Stack

• Frontend: React 19, Vite, Tailwind CSS (v4), Motion (for animations), Lucide React (icons).
• Backend: Node.js, Express.js, TypeScript.
• AI Engine: Google Gen AI SDK (@google/genai), using the Gemini model.
• Build/Deploy: esbuild for compiling the backend, bundled for containerised Cloud Run deployments.

Outcomes and Trust Mechanisms

• Standardised Output: Generates a uniform "Assessment Pack" that judges all AI projects on the same scorecard.
• Risk Surfacing: Highlights technical dependencies, compliance red flags, and "when to stop" thresholds early.
• Actionable Next Steps: Delivers a clear "Go / No-Go / Pilot / Options" verdict with explicit justifications.
• No Hallucinated Infrastructure: Forces the model to operate strictly within the provided bounds, explicitly listing assumptions when estimating missing data.

Future Roadmap

• Multi-Agent Orchestration: Transition from a single-prompt MoE to a true multi-agent system where dedicated functional agents (e.g., a Legal Agent, a Cloud FinOps Agent) negotiate the final assessment.
• Integration with Enterprise Architecture Tools: Export assessments directly to Jira, Confluence, or ServiceNow.
• Historical Accuracy Tracking: Allow users to update projected vs. actual costs/latency to continuously train and calibrate the estimator's baseline assumptions.
• Expanded Regulatory Frameworks: Implement specific knowledge bases for the EU AI Act, HIPAA, SOC 2, and specialised regional privacy laws tailored to specific industry verticals (including Australian Privacy Principles - APPs).