Proofy
Proofy helps journalists, analysts, researchers, investors, and meeting participants verify factual claims made during conversations, startup pitches, or recordings. Checking spoken claims manually requires identifying each claim, finding credible sources, looking for counterevidence, and comparing the findings. This interrupts conversations and makes misleading statements difficult to challenge quickly. Proofy transcribes live speech or uploaded audio/video, extracts verifiable claims, and searches for supporting and contradicting evidence. Each completed check includes a verdict, source excerpts, an explanation, and evidence confidence. Users can also provide PDF, TXT, and CSV documents as additional sources.
Preliminary score
- Value for the user16.5 / 25
A specific problem and a clear outcome from solving it.
- Prototype and use of AI22.5 / 30
A working core scenario, and what AI actually contributes to the result.
- Quality testing12.5 / 20
Tests, examples of failures, and comparison with the current approach.
- Feasibility9.5 / 15
Data requirements, running costs and a clear next step.
- Originality6.0 / 10
A meaningful difference from common solutions.
What the judges wrote
- Claude Opus 5.5Preliminary64/100
- Strengths
- Your prototype covers a real end-to-end pipeline: live or uploaded speech goes through Speechmatics transcription, Gemini claim extraction, Bright Data web research and a deliberate counter-evidence stage, ending in a four-way verdict with confidence, and it is deployed. Your engineering discipline stands out. You report 70 automated tests covering quote validation, source independence and verdict calculations, and you dropped an optimization because it produced invalid quotations. Separating provider failures such as 429 and 503 errors from factual verdicts, and defining confidence as evidence strength rather than truth probability, are thoughtful choices that build trust.
- To improve
- Your testing is strong on software correctness but has no accuracy evaluation yet. Run even 30–50 labeled claims, report how many verdicts were right, and show the failures. Your own deck notes reveal a verdict-logic bug: 'Insufficient evidence' was shown with 2 refuting and 0 supporting sources. Either show it fixed or state it openly. For the Enterprise track, pick one first customer, such as a newsroom or an investor team, and name an adoption metric: time saved per claim versus manual checking, or precision of flagged claims. Also quantify the cost of one Quick and one Detailed check in speech-to-text, LLM and search calls.
- GPT-6 AstraPreliminary70/100
- Strengths
- Your prototype connects speech transcription, claim extraction, and supporting/counter-evidence research into an inspectable workflow, with AI contributing directly to the result. Your reported 70-test suite addresses relevant risks, including invalid quotations, source independence, and provider failures. Separating technical failures from factual verdicts and defining confidence as evidence strength are sensible safeguards.
- To improve
- Choose one initial enterprise workflow and define an adoption metric, such as verification time saved at an acceptable false-verdict rate. Reconcile the verdict mismatch noted on slide 9 and demonstrate the final logic end to end. Add a labeled-claim evaluation covering Azerbaijani input, ambiguous claims, and unreliable sources, with a manual-checking baseline. Measure transcription, model, and search costs per session to validate whether your $15 unlimited tier is sustainable.
The team
Nijat Orujov
developer
Toghrul Allahverdiyev
data
Mahammad PiriyevLead
developer