Chaque idée devait montrer un acheteur prêt à payer, des preuves tirées de la discussion et un vrai manque, puis survivre à une red team chargée de l'abattre.
Articles passés en revue
25
Lus en détail
6
Idées proposées
12
Retenues
2/12
2 idées retenues
| # | Note | Idée | Type | Catégorie | Douleur | Payer | Timing | Manque | Défense | Faisable |
|---|---|---|---|---|---|---|---|---|---|---|
| 1 | 6.6 | Plateforme évolutive d'examens oraux pour universitésLogiciel et processus permettant aux départements de faire passer à grande échelle de courts examens oraux structurés (soutenances), le format d'évaluation que l'IA ne peut pas imiter. | Venture | Éducation | 7 | 6 | 8 | 7 | 6 | 6 |
| 2 | 6.0 | Kit de démarrage on-prem pour modèles à poids ouverts, en serviceConseil productisé et outillage qui font passer des entreprises de taille moyenne de zéro à un déploiement on-prem de LLM ouverts opérationnel en quelques semaines. | Indie | Infrastructure IA | 6 | 6 | 7 | 6 | 4 | 7 |
Écartées · 10 · sous la ligne de 6.0
| 5.9 | Inline PR annotations instead of review reports | Scored 5.9, under the 6.0 bar. |
| 5.9 | Pre-deploy blast radius and on-call router | The supporting evidence is essentially one engineer who hand-built an internal tool, and stitching diff-reach analysis to PagerDuty ownership is a hard PagerDuty-Startup feature rather than a defensible venture company. |
| 5.9 | LLM provider policy and privacy posture monitor | Scored 5.9, under the 6.0 bar. |
| 5.3 | Continuous model swap advisor for LLM workloads | This is the same product as the token-spend router candidate (re-evaluate models against your prompts, route to cheapest acceptable), and OpenRouter plus Artificial Analysis already supply most of it free. |
| 5.2 | Contradiction-aware memory store for LLM agents | Contradiction detection and decay are roadmap features Mem0 and Zep can add within quarters, and agent builders mostly stay on provider-native memory unless the pain forces migration, making this a feature underlying platforms will absorb. |
Voir 5 idées écartées de plus
| 5.2 | In-person exam workflow for handwritten STEM assessment | Gradescope and Crowdmark already do scan-and-rubric grading of handwritten STEM exams at exactly this price point, leaving only marginal workflow tailwinds and no durable gap an incumbent can't close in one release. |
| 5.0 | Token-spend router across frontier and open models | Cheapest-model-acceptable routing is the flagship feature set of OpenRouter, LiteLLM, and Portkey, and the 'quality-eval' differentiator is a feature Martian and the eval tools already bolt on — this is not a company. |
| 4.8 | Benchmark suite for context management strategies | The demand signal is one commenter wishing for this benchmark, the natural outcome is a published comparative report (freely consumed) or a Braintrust eval template, and a $5k one-time report to a handful of platform teams is not a company. |
| 4.7 | Offline skin lesion triage while awaiting dermatology | Too weak on moat (3/10). |
| 4.2 | Turn-key narrow classifier builder for image triage | AutoML and Roboflow already train narrow classifiers from small labeled sets, and 'here's proof it beats GPT on your task' is a sales demo, not a product companies will pay $800/mo for. |