ProofPilot

Inspiration

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for ProofPilot

Inspiration

Important decisions are often made from scattered tabs, long reports, and conflicting claims. The hard part is not finding more information—it is knowing what to trust, what is missing, and how each source supports the final recommendation. We built ProofPilot to turn that messy process into an auditable conversation.

What it does

ProofPilot is an AI research agent that accepts a decision question and a collection of sources. It breaks the question into research tasks, extracts relevant evidence, compares conflicting claims, and produces a concise decision brief. Every recommendation is linked to supporting evidence, while uncertainty and unresolved gaps are shown explicitly.

A typical workflow is:

1. Ask a real-world question.
2. Add webpages, PDFs, notes, or structured data.
3. Let the agent investigate in parallel and challenge weak evidence.
4. Review a structured brief with citations, confidence levels, trade-offs, and recommended next actions.

How we built it

The prototype uses the OpenAI API for planning, tool selection, evidence extraction, synthesis, and structured outputs. A lightweight orchestration layer keeps each research step traceable. Sources are normalized into evidence records, then passed through a verifier that checks whether every major claim has direct support before the final brief is generated.

The interface is designed around evidence cards rather than a generic chat transcript. Users can inspect which source supports a claim, see contradictions, and reopen the original context without losing the decision narrative.

Challenges we ran into

The biggest challenge was balancing helpful synthesis with epistemic honesty. Models can write persuasive prose even when the evidence is thin, so we treated citation coverage, contradiction detection, and visible uncertainty as core product features—not optional polish. We also had to keep long research sessions understandable without overwhelming the user.

Accomplishments that we're proud of

• A decision-first workflow instead of another search or chat interface
• Claim-level traceability from recommendation back to source evidence
• Explicit handling of contradictions, missing evidence, and confidence
• Structured outputs that can be reused in reports, tickets, or team discussions

What we learned

Trust comes from interaction design as much as model quality. Showing why an answer was produced, where it may be wrong, and what evidence would change it makes an AI agent far more useful for consequential work.

What's next for ProofPilot

Next, we want to add collaborative review, reusable research playbooks, live source monitoring, and domain-specific verification tools. The long-term goal is to make evidence-backed decision making as fast and natural as asking a question.

Built With

• openai-api