Ideator
An open-source AI product decision assistant based on knowledge graphs. It builds a competitor knowledge graph through multi-path search and uses a Gemma agent to reason over the graph, outputting build briefs with sources.
Timeline (last 90 days)
Developer Ronak Sarda submitted a project called Ideator in the second week of the Hacktoberfest open-source AI challenge.
Ideator gathers information by running three live searches in parallel: Google to find competitors, Google Play to find app names and pricing, and Reddit to find user complaints.
Ideator uses the FalkorDB graph database to store a separate knowledge graph for each session, with nodes including companies, prices, features, and complaints.
Ideator's final output is a five-box build brief containing five slots: who pays, what pain it solves, why it wins, pricing, and the first version.
Ideator's pricing slot is automatically filled by the TabPFN model, the pain point slot is automatically filled with evidence, and the remaining three are decided by the user.
Ideator uses Cypher queries to perform multi-hop reasoning on the knowledge graph, looking for feature points that users of multiple competitors are complaining about but that the user's idea has not yet addressed.
Ideator uses gemma-4-26b-a4b-it and then gemma-4-31b-it to run the agent loop and write the report.
When the hosted Gemma endpoint fails, Ideator's short prompts fall back to the gemma2:2b model on local Ollama.
Ideator's TabPFN pricing tool trains a regressor on the spot on each call, and price estimates become more accurate as the number of uses increases.
The author found that putting process constraints at the code layer (such as REQUIRED in tools.py) is more reliable than at the prompt layer, because the model sometimes skips the steps required by the prompt and answers directly.
3 more timeline events
All Facts (13)
Several of the author's college friends said after trying Ideator that they would each be willing to pay about 2,500 rupees to use it, but no actual payment has occurred yet.
50%UnverifiedIdeator gathers information by running three live searches in parallel: Google to find competitors, Google Play to find app names and pricing, and Reddit to find user complaints.
50%UnverifiedIdeator uses the FalkorDB graph database to store a separate knowledge graph for each session, with nodes including companies, prices, features, and complaints.
50%UnverifiedIdeator's final output is a five-box build brief containing five slots: who pays, what pain it solves, why it wins, pricing, and the first version.
50%UnverifiedIdeator's pricing slot is automatically filled by the TabPFN model, the pain point slot is automatically filled with evidence, and the remaining three are decided by the user.
50%UnverifiedIdeator uses Cypher queries to perform multi-hop reasoning on the knowledge graph, looking for feature points that users of multiple competitors are complaining about but that the user's idea has not yet addressed.
50%UnverifiedIdeator uses gemma-4-26b-a4b-it and then gemma-4-31b-it to run the agent loop and write the report.
50%UnverifiedWhen the hosted Gemma endpoint fails, Ideator's short prompts fall back to the gemma2:2b model on local Ollama.
50%UnverifiedIdeator's TabPFN pricing tool trains a regressor on the spot on each call, and price estimates become more accurate as the number of uses increases.
50%UnverifiedThe author found that putting process constraints at the code layer (such as REQUIRED in tools.py) is more reliable than at the prompt layer, because the model sometimes skips the steps required by the prompt and answers directly.
50%UnverifiedIdeator takes about one to two and a half minutes to generate a complete analysis of a new idea, and follow-up questions are faster.
50%UnverifiedIdeator's search still goes through SerpApi and its price prediction goes through Prior Labs, so full localization is currently a direction rather than the status quo.
50%UnverifiedDeveloper Ronak Sarda submitted a project called Ideator in the second week of the Hacktoberfest open-source AI challenge.
50%