Innovation sketches

Preseed Canvas — simplifying investment decisions

The Preseed Canvas is a structured way to assess potential and progress of early-stage startups. It aims at all aspects relevant for a decisions beyond interaction with the team.

• Preseed Canvas: Whitepaper on the concept of Preseed Canvas and its criteria and scores (updated: 15 Dec. 2024).
• Canvas App allows filling in 8 different canvas types including Preseed Canvas -- without sharing data with a server (updated: 10 Jun. 2024).
• Pitch2Canvas is an upcoming application of the Preseed Canvas to perform automated analysis and scoring of pitch decks and other documents (updated: 15 Dec. 2024).

Motivation: We are seeing 100 decks for a single investment...

Semantic RAG with LLMs — towards a local knowledge engine

Current LLMs cannot reason, current "agentic" systems are often not reliable because of this, which led to rollbacks of various high-profile AI initiatives. With semantic RAG we target cutting-edge functionality that sets out to solve a few challenges with today's LLM and RAG systems, especially in business contexts: Explainability and hallucination busting via a semantic reasoning layer ("neurosymbolic approach"), the notion of answer confidence, dynamic objective-based and interactive research strategies, and various forms of annotations for organization, feedback and model improvement. This project should replace previous work in Pitch2Canvas.

• Ragster: A modular, extensible, privacy-first system for conducting deep research using local or remote LLMs and a number of RAG sources that is the basis of the semantic RAG approach.
• Semi-formal languages (SFLs) are a continuum of knowledge representations that range between unstructured data (including text token sequences) on one hand and formal languages (including ontologies) on the other by providing a flexible framework with varying degrees of formality and factual confidence. We posit that SFLs can be a key building block for neurosymbolic systems that is easy to apply, both for extracting SFLs from unstructured data and for making LLM inference (more) explainable and add true deductive reasoning.

Motivation: LLMs have hit a wall in reliability and non-trivial deductive reasoning tasks. A solution may build on Gregor's PhD work.

Creating Data Products — unlocking value through AI and semantics

Data products and data-driven products are products where data drives — or is — the core value proposition. They transform raw information into actionable insights, automated decisions, or intelligent services that create measurable business value. The rise of AI, particularly large language models (LLMs), has dramatically expanded the possibilities for data products. Semantic relations between data are the key to establishing consistency and predictable reasoning. We discuss implications, opportunities and challenges beyond the hype.

• Data Product Management: Describes my view on data products and what it takes to create, develop and optimize them. Establishes the role of AI and semantic processing in future data products.
• From data to knowledge products: Data are at the core of today’s processes. So is knowledge not captured in the data. We explore a way how a new type of product can marry both.
• A typology of AI and data products: Proposes a pragmatic way of classifying AI and data products and predicts implications on investment valuation.

Motivation: Gregor has worked 15 years on data and AI products. This is an application of the Semantic RAG work above.

Pipeline Processes — optimizing human+machine decisioning

Pipeline processes, or short: pipelines, are multi-step decision processes that filter a set of candidate items into a result set, given appropriate criteria. This captures the dynamics of complex decision making under resource constraints, which is core to fields as wide as angel investing, recruiting, sales (esp. B2B), product management and, as an extension, multi-agent systems. Analyzing today’s processes as pipelines allows a structured approach to improve efficiency and effectiveness through optimizing the balance and sequence of human and machine decisions, the decision criteria themselves, consistent valuation of live pipelines and best-practice reuse between application areas.

• Pipeline Processes — an overview: Introduces pipelines as a structure and links it to the extensive but disparate work in various fields of application and research. The result is a tool to help transform and optimize real-world pipelines with the right data and the right decisions with the right AI augmentation (planned: early 2025).
• Pipeline Processes — a model: The article gives an overview of a mathematical formulation and proposes a gradient-based solution to optimizing pipelines (updated 22 Nov. 2024).
• Sales Pipeline Analysis: A study analyzing a real-world sales funnel as a pipeline process (planned: early 2025).
• B2B Market Making: Shows how many B2B BD and sales pipelines are extremely inefficient and proposes a solution based on an AI-based communication channel that directly links into existing tools (planned: early 2025).

Motivation: A largely unrepresented area in AI and CRM where some innovative approaches could drive significant improvements.

Project "Lakeshore" — from data mess to data mesh

Data are at the root of any digital or AI business. Unfortunately, making data sources ready for digital business often turns out to be an organizational and technical nightmare. Data seem like the "dark matter" of AI in that they exist but are very difficult to get a grip on. Project "Lakeshore" aims at a knowledge graph-based platform that transforms scattered enterprise data on-premises and in SaaS systems into a unified, searchable, and AI-ready resource. It can serve as a semantic enterprise search engine and as a platform for AI and data management initiatives — in other words as an enabler for digital transformation, especially for SME companies.

• Lakeshore architecture: Overall concept and architectural setup for the Lakeshore system centered around a knowledge graph as semantic overlay, a structured indexing mechanism and a fine-grained access control mechanism.
• Hybrid semantic indexing: Indexing data sources into semantic structures, using probabilistic approaches, LLMs and ontology learning.
• Hybrid semantic retrieval: Combining classic retrieval systems with semantic and conversational approaches, using probabilistic, LLMs, RAG and knowledge graphs. Think of it as an enterprise "perplexity.ai" with structured reasoning on top.
• Privacy-preserving access control: Outlines the structure and algorithm to ensure confidentiality for federated source systems, private access spheres and levels; considers semi-private use cases like federated learning.
• Data spaces and composite learning: Bringing together mutually access-restricted data for federated access. This extends Federated Learning.

Motivation: Gregor has seen SaaS application sprawl in many organizations. Integrating their information has grown in importance when data are needed to train and finetune AI.

Project "Unleash" — de-risking to drive innovation

Angel investment and Venture Capital portfolio returns are based on long-tail distributions: Outsized returns of very few investments drive the success of the whole portfolio. For founders, there's only a slim chance to become commercially successful. For this reason, many innovators do not become startup founders and many investors choose less risky asset classes. Project "Unleash" aims at reducing risk for both parties.

• Pareto Unskewed: Analyzes the return distributions of venture portfolios and discusses pool-based de-risking strategies. Proposes the concept of a "reverse insurance".
• VC Continuation Fund: Proposes the concept of a de-risked funding mechanism for angel networks that may make the asset class also more attractive for individuals and family offices.

Motivation: We are interested in making startups more attractive -- both as a career choice and as an asset class for small investors. This requires a deeper look at the risk profile and potentially some ways to adjust it.

"How cool is ... ?!" — analyses and opinions

Discussing miscellaneous technical, business and leadership topics, this series highlights a couple of insights and opinions.

• "How cool are ... "static" webapps?!": Client-only, "static" web apps are a very secure and scalable way of delivering content. This note gives an overview of the pros and cons. For many use cases, it is an option often overseen.
• "How cool is ... the future of Product Management?!": We look at some AI-induced developments across industries and posit that future PMs will have a widened role and be much more central to business success.
• "How cool are ... Causal Loop Diagrams?!": CLDs are a great way to describe complex systems and strategies. We introduce the method and propose an AI-based way to get the most out of them.
• "How cool is ... cloud-first security?!": Doing cloud-first seems the obvious thing for a new company. However, cybersecurity auditors tend to know shops centered around on-prem. This note shows how to establish a good security posture without even an intranet.
• "How cool are ... the first 100 days as a CTO?!": I did this journey several times and there's a pattern I am happy to share here.