Sage Decision Intelligence

01 Main Header Page
06 Side Drawer on Evidences, Recommendations
05 Complete Analysis
04 Recommendation main page
02 Files / Artefacts upload section
03 Files processing and AI analysis in action
04.1 Recommendation main page1

Inspiration

High-stakes enterprise initiatives are often approved using fragmented information spread across project charters, requirements documents, meeting notes, delivery plans, and governance records.

Executives are then asked a deceptively simple question:

Should this initiative proceed to execution?

The problem is that hidden dependencies, missing owners, regulatory gaps, weak assumptions, and delivery risks are often discovered only after significant money and time have already been committed.

We built Sage Decision Intelligence to provide leaders with an intelligent second opinion before execution begins.

What it does

Sage Decision Intelligence transforms project documents into an executive-ready decision assessment.

Users upload initiative artifacts and provide brief business context. The application then uses GPT-5.6 to produce:

An overall recommendation:
- Proceed
- Proceed with Conditions
- Do Not Proceed Yet
A Project Health Score
An assessment confidence level
A concise executive summary
Prioritized Decision Signals
Regulatory, timeline, and financial exposure
Evidence supporting each finding
Recommended next actions and suggested owners

Each Decision Signal can be opened in a detailed side panel showing why it matters, the supporting evidence, and the recommended mitigation.

The product is intentionally not a chatbot. It is a focused, dashboard-first decision experience designed for enterprise executives, transformation leaders, program directors, product leaders, and PMOs.

How we built it

We designed the product around one clear workflow:

Upload documents → Analyze evidence → Review executive recommendation → Investigate Decision Signals → Take action

The application was built with:

Next.js and React
TypeScript
Tailwind CSS
shadcn/ui
OpenAI Responses API
GPT-5.6
Strict Structured Outputs using JSON Schema
OpenAI Node SDK
react-dropzone
Mammoth for DOCX extraction
PDF text extraction
Session-based state for the hackathon MVP

The AI layer uses a single server-side analysis workflow rather than a complex multi-agent system.

Submitted materials are treated as untrusted reference data. The prompt explicitly prevents instructions embedded inside uploaded documents from overriding the system behavior.

GPT-5.6 returns a strictly validated AnalysisResult, allowing the same typed contract to power both the API and the executive dashboard.

Challenges we faced

Producing reliable executive outputs

A generic model response was not sufficient. The output needed to be concise, evidence-backed, prioritized, and immediately usable by senior decision-makers.

We solved this using a strict JSON Schema, runtime validation, explicit evidence rules, and a carefully scoped enterprise transformation prompt.

Avoiding fabricated or repetitive findings

An early schema required a fixed number of Decision Signals. With sparse documents, this encouraged repeated findings.

We changed the contract to allow zero to five evidence-supported signals and required overlapping observations to be consolidated into one distinct finding.

Preserving trust

Enterprise documents may contain sensitive information or even prompt-injection attempts. We kept all model calls server-side, disabled response storage for the analysis call, avoided logging document contents, and explicitly treated uploaded content as untrusted evidence rather than instructions.

Designing for executives rather than AI enthusiasts

It was tempting to create a conversational interface. Instead, we focused on the decision executives actually need to make and surfaced the recommendation, health score, exposure, evidence, and next actions in a calm, dashboard-first experience.

Accomplishments that we are proud of

Built a polished enterprise decision dashboard rather than a generic chatbot
Integrated GPT-5.6 through the OpenAI Responses API
Implemented strict Structured Outputs and runtime validation
Created evidence-backed Decision Signals with confidence scores
Added distinct zero-signal and low-evidence behavior
Built a detailed signal investigation drawer
Added real document extraction for common enterprise file formats
Created an end-to-end workflow from uploaded artifacts to live executive assessment
Designed a product experience that feels credible for enterprise use while remaining achievable as a hackathon MVP