AmAmAm
Students, researchers, and knowledge workers regularly struggle to navigate information-heavy websites, including articles, documentation, discussion forums, and online platforms. Finding relevant information requires endless scrolling, repetitive keyword searches, and manually filtering irrelevant content. Traditional browser search relies on exact keywords, while standalone AI chatbots can explain information but cannot directly connect their answers to specific elements on the webpage or visually reorganize its content. Users need a faster, more intuitive way to find, understand, and interact with information without leaving the website they are browsing.
Preliminary score
- Value for the user14.5 / 25
A specific problem and a clear outcome from solving it.
- Prototype and use of AI23.0 / 30
A working core scenario, and what AI actually contributes to the result.
- Quality testing13.5 / 20
Tests, examples of failures, and comparison with the current approach.
- Feasibility8.0 / 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 5.5Preliminary60/100
- Strengths
- Your architecture keeps a clear safety boundary: the model only selects targets through document-local IDs, while deterministic code validates them, rejects stale snapshots, and applies reversible effects. Your 185 automated tests (unit, integration and browser E2E), including safety tests for invented IDs, wrong excerpts and unsolicited actions, show real engineering work. You were honest about limits, naming Instagram and dynamic, image-based sites as weak spots and stating that accuracy, latency and cost are not yet measured.
- To improve
- You entered the Enterprise track but describe a general audience of students and researchers, with no named company workflow and no adoption metric. Choose one concrete case, such as support agents searching an internal documentation portal, and measure time-to-answer against today's process. Your tests check that the code behaves correctly, not that the AI is accurate. Build a labelled set of queries across real sites and report grounding precision, action correctness, latency and cost per query using gpt-4.1-mini and embeddings.
- GPT-6 AstraPreliminary70/100
- Strengths
- Your prototype connects grounded answers to real webpage elements and reversible actions, providing more than a standalone chat interface. Your separation of AI target selection from deterministic execution, including stale-snapshot and unsolicited-action checks, is a strong engineering choice. Your reported 185-test suite covers browser behavior, grounding, and multilingual retrieval, while your disclosures clearly acknowledge dynamic-site failures and unmeasured performance.
- To improve
- Choose one enterprise documentation workflow and measure task completion time and answer correctness against manual browsing and Ctrl+F. Separate real-model evaluations from deterministic or mocked tests, and report target-selection accuracy and latency across representative websites, including dynamic pages. Quantify API cost per task and define a practical deployment path beyond a locally operated backend. Clarify which webpage content reaches the model provider: local chat storage does not justify the broader claim that data stays in the browser.
The team
Ali FarajovLead
product
Islam Hasanli
business
Kenan Nuraliyev
data