At Rise of AI Summit 2020 in Berlin, moderator Jasper Masemann (Partner, HV Capital) hosts a two-and-a-half-hour startup demo session in which founders and executives from ten German AI companies pitch their products and take short audience Q&A. Nicole Büttner presents Merantix Labs as an external machine learning department for corporate NLP, computer vision and forecasting projects. Stefan (neurocat) presents automated robustness, safety and explainability testing for deployed AI, including an 'attack generator' for stress-testing models. Alexander Stojanovic presents AVA, a data-standardization platform for security, safety and risk information used by organizations such as London's police and DHL. Micropsi Industries and Wandelbots present software for training industrial robots from camera and sensor data without conventional programming. i2x, Solvemate and Cognigy present conversational AI for customer service, ranging from real-time agent coaching to self-service chatbots and voice automation. A privacy-technology founder presents 'AI for Privacy', a case-based tool for automating GDPR risk assessments, and Clemens (Sia Search) presents a metadata-search layer for large unstructured computer-vision datasets. Each pitch covers the product, target customers and a short Q&A on differentiation and roadmap.
Key takeaways
- Merantix Labs positions itself as an outsourced machine learning department, citing legal, pharma and manufacturing case studies rather than a single vertical product.
- neurocat automates AI robustness and explainability testing with an 'attack generator' that runs far more stress tests than manual methods.
- AVA normalizes security, safety and risk data from many sources into one common, patented standard used by London police and DHL's global security team.
- Wandelbots and micropsi Industries let non-programmers teach industrial robots by demonstration or camera-driven control instead of conventional coding.
- i2x, Solvemate and Cognigy each automate customer service differently: live agent coaching, self-service deflection, and no-code voice/text virtual agents.
Detailed notes
Merantix Labs: machine learning as an external service
Nicole Büttner, CEO and co-founder of Merantix Labs, describes the company as an external machine learning department that helps corporate partners with concrete NLP, computer vision and forecasting problems rather than selling a single vertical product. Named examples include automating legal compliance review for a 'magic circle' law firm, classifying customer product reviews, flagging new pharmaceutical side-effect reports, building skin-cancer classifiers from dermatology and histopathology images, and helping manufacturers translate product sketches into production specifications. She positions Merantix Labs against traditional consultancies by pointing to its engineering ('tech DNA') rather than presentation-building background, and describes an 'AI canvas', developed with the University of St. Gallen, structured around business purpose, organization and technology/data pillars to help companies move machine learning from prototypes to sustainable, scaled deployments. She also describes building cross-company 'data ecosystems' in specific industries, such as automotive supply chains, so partners can create shared value from linked data.
neurocat: robustness and explainability testing for deployed AI
Stefan, a co-founder of neurocat, presents a self-funded, bootstrapped company built to address robustness, safety, security and explainability gaps in deployed AI, motivated by real-world failures such as chatbots developing unintended behavior and a UK court case over an unexplainable trading AI. neurocat's testbed can automatically generate very large numbers of adversarial 'stress tests' (described as billions, versus roughly 50 to 200 in earlier manual approaches) through an 'attack generator', and pairs this with an explainability tool offering local and global explanations. The company works across the automotive certification and insurance chain and cites Volkswagen and BMW as customers, plus new projects in telecom and explainable AI consulting, and says its methods are input- and industry-agnostic, applicable to speech, vision or other model types.
AVA: a common data standard for security and risk
Alexander Stojanovic presents AVA, founded after he witnessed a car-ramming attack near a conference in Austin in 2014 and concluded that available sensor and social data could have warned people in advance if it had been connected. AVA ingests incident data (crime, fire, accidents, disease outbreaks, terrorism, social and political risk) from customer, public and personal-ownership sources and normalizes it into one common, patented data structure with a unique ID per event, enabling organizations to share risk information without sharing underlying data. Stated customers include the police of London and London City Hall, public transport operators, and DHL, whose global security operations center uses AVA to monitor risks to its logistics network worldwide. Stojanovic argues the value lies less in individual AI techniques than in establishing AVA's common standard as infrastructure other systems build on.
Robotics: micropsi Industries and Wandelbots
Speaking virtually, micropsi Industries' presenter describes plug-and-play control software that drives industrial robots directly from camera and force-torque sensor input for unstructured tasks such as picking and reorienting wine glasses on a conveyor, work that traditional pre-programmed automation cannot handle. He cites falling cloud training times (from about 16 hours a year earlier to roughly two and a half hours, with a projected half hour) and says a customer can go live within roughly a week once physical setup is in place. Christian, CEO of Wandelbots, presents software that lets non-programmers teach industrial robots by demonstration with a tracked handheld device, generating code for multiple robot brands (KUKA, ABB, Universal Robots). He cites a Volkswagen end-of-line inspection use case needing only 20 to 50 reference images per check point to reach about 99 percent accuracy, and a demo in which Microsoft CEO Satya Nadella remotely taught a robot in Germany a sealing process from Redmond. He says less than one percent of manufacturing jobs that could be automated currently are.
Conversational AI for customer service: i2x, Solvemate, Cognigy
Michael, founder of i2x, presents real-time speech-to-text analysis and live coaching for call center agents, delivered to the agent's screen in under half a second per cycle, alongside customer-satisfaction and emotional-tone analytics; he cites 10 to 30 percent conversion increases and about 50 percent time savings for team leads, and frames the product as agent augmentation rather than replacement. Eric, founder of Solvemate, presents a self-service chatbot platform built on what he calls 'dynamic decision trees' rather than pure NLP, aimed at deflecting routine requests for companies with 10 to 200 service agents; he says setup can go live the same day and cites Gartner's AI hype cycle to argue chatbots are now past the hype phase. Silas, from Cognigy, presents a low-code/no-code platform for building voice and text virtual agents across more than 30 channels, offered as SaaS or on-premise (on-premise being the preference of GDPR-sensitive customers such as insurers and German companies), citing Lufthansa as a client and reporting around 30 percent call deflection.
AI for Privacy and Sia Search: governance and data tooling
A privacy-technology founder, a lawyer by training, describes building the 'Privacy Suite' after his team spent 70 hours on a single GDPR data protection impact assessment; the tool now automates risk scoring and, through 'AI for Privacy', uses case-based reasoning across anonymized processing activities to suggest risk mitigations, funded in part by an innovation grant from the German state of Hesse. Clemens, from Sia Search (CSH), presents a metadata layer for large unstructured computer-vision datasets, arguing that engineers can spend up to three-quarters of their time on manual data handling; the product uses interval-based metadata matching, which he says is roughly three orders of magnitude faster to query than conventional per-frame databases, and is offered free for academic use on the nuScenes dataset.
Quotes
We're basically your external machine learning department for companies.
Nicole Büttner
There is no minute passing without somewhere in the world stuff happening.
Alexander Stojanovic
The best customer service is no customer service.
Eric (Solvemate)
Questions answered in this talk
What does Merantix Labs offer companies building machine learning products?
According to Nicole Büttner at Rise of AI Summit 2020, Merantix Labs works as an external machine learning department, building bespoke NLP, computer vision and forecasting solutions for partners in fields such as legal compliance, pharmaceutical safety monitoring and manufacturing, and helps organizations structure their path from prototypes to scaled, sustainable deployment.
How does neurocat test the robustness of AI systems before deployment?
According to neurocat's co-founder at Rise of AI Summit 2020, the company's 'attack generator' automatically produces very large numbers of adversarial stress tests against a model, far more than manual testing could achieve, and pairs this with explainability tooling to give local and global explanations of model decisions, used in the automotive certification and insurance chain.
How do Wandelbots and micropsi Industries let non-programmers control industrial robots?
According to the presenters from Wandelbots and micropsi Industries at Rise of AI Summit 2020, their software lets users teach a robot by demonstration with a handheld tracked device, or drives the robot directly from live camera and sensor input, generating robot-specific control code without conventional text-based programming.