# A demand signal system for finding startup opportunities

I wanted a way to find startup ideas that does not depend on taste or on a clever prompt. I asked an AI to design the system as a research problem: where the evidence comes from, how it is ranked, and how an idea is proven or killed.

This is a design, not a product I have built or a list of opportunities I found.

## What I found
- The proposed system keeps a source record for every claim. It connects collection, search, extraction and market sizing to experiments that test whether people want the result.
- Search spikes and social buzz are weak hints. The system has to model pain, workaround cost, repetition, who the buyer is, budget signals and willingness to pay, which is why it mixes discussions, issue trackers, reviews, job posts and tenders.
- The study proposes joining demand discovery and experiment-based validation. Its comparison of existing tools is preliminary. Google Trends shows search interest, Semrush and similar tools show attention, and general AI platforms help build pipelines. I did not test those products.
- Retrieval should be hybrid and multilingual: keyword search, dense embeddings and a reranker, with a model that covers more than 100 languages so Russian and English evidence sit in one index.
- Every idea gets a score with eight stated factors, for example pain severity at 20%, frequency at 15% and willingness to pay at 15%. The weights are a proposal. This weighted score is a priority index, not a probability. A separate probability model would need calibration against defined outcomes in held-out data. For example, seven successes out of ten illustrate what 70% means, but ten cases would not establish calibration.
- Validation is a ladder from cheap to costly: search trends, stated pain, landing pages and outreach, a manual pilot, then proper tests. Variance reduction and difference-in-differences can help estimate effects when the study design supports their assumptions; neither guarantees that a lucky week will not be mistaken for evidence.
- Research can be broad, but acting needs explicit approval: spending money, contacting people or changing live assets. The study estimates a first serious version at about 41 to 62 person-months.

## How it was made
The study was produced on 17 May 2026 with ChatGPT Deep Research. It draws on published work on retrieval, topic clustering, weak supervision and experiment design, plus documentation of existing tools. The AI researched and drafted. I asked the question. Tool features and platform rules were current at that date.

## Sources
- [Lewis et al., Retrieval-augmented generation for knowledge-intensive NLP](https://arxiv.org/abs/2005.11401)
- [Santhanam et al., ColBERTv2](https://arxiv.org/abs/2112.01488)
- [Chen et al., BGE M3 embedding](https://arxiv.org/abs/2402.03216)
- [Grootendorst, BERTopic](https://arxiv.org/abs/2203.05794)
- [Ratner et al., Snorkel](https://arxiv.org/abs/1711.10160)
- [Ash et al., BADGE batch active learning](https://arxiv.org/abs/1906.03671)
- [Chernozhukov et al., double debiased machine learning](https://arxiv.org/abs/1608.00060)
- [Wager and Athey, causal forests](https://arxiv.org/abs/1510.04342)
- [Gebru et al., Datasheets for datasets](https://arxiv.org/abs/1803.09010)
- [Mitchell et al., Model cards for model reporting](https://arxiv.org/abs/1810.03993)
