/ Artificial Intelligence

Part 3 - Using AI to Consistently Deliver Successful Projects and Build Teams

Matching the Right Resources

The rich amount of task-level metadata available about a given project allows us to build recommendation systems to a far more granular level than ever before possible.

A properly scoped-out project is the starting point for our core services, around which we provide the necessary connections to the world’s best freelance talent.

Once a project is planned, we are automatically able to support many decisions that were previously out of reach without a dedicated project manager guiding and aligning the setup of the required mix of talent and resources to deliver the project.

Setting up a project team involves decisions about:

  • What qualification criteria will enable us to hire the best fit to deliver on a given scope of work?
  • What factors in personality and work style will positively influence team dynamics and lead to an overall optimized project delivery?
  • How do you select a team of people from a public talent pool so as to maximize team fit and minimize friction?

Matching the Right Talent

This introduces a scenario of multi-level optimization, where you need to maximize the likelihood of acquiring the best talent to work on the most important aspects of the project, while working within the constraints of of insuring compatibility in communication and work styles and (inter)personal values.

Project managers could be building a team from scratch around a new project, or they could be looking to scale an existing team quickly with new experts drawn from global markets.

This introduces many friction points concerning qualifications and skill-sets, and securing objective measures of trust and dependability to insure against intentionally fraudulent activity.

Recommendation

Modern recommendation Engines use a process of vector representation of the input data to represent the complex non-linear relationship between the features, and a special distance minimization function to identify a close alignment between the analyzed objects.

We are working with Daniel Shapiro, PhD, who wrote his thesis on the topic of Recommender Systems to tackle the complex, hierarchical nature of the recommendation task we are implementing.
quote_daniel

To inform the recommendations we maximize data quality and trust within the ecosystem. Together with our Advisor Fabian Vogelsteller (originator of the ECR20 Standard for Smart Contracts on the Ethereum blockchain), we are working to implement the ERC725 protocol: A new standard for personal data, enabling portability of blockchain verified professional qualifications and background checks.

quote_fabian

Together with our partners we are providing an integrated process to help Experts quickly port their off-chain data from existing platforms onto the chain.

Past success tells a great story

As we have learned from our product validation, former successful collaboration is the most important qualification for project leaders looking to expand their team. This is due to a “chain of trust”, which is normally established manually and off-line, but can now be scaled to a global network of talent via DREAM.

From assessing skills, to verifying talent, project endorsements, reputation building, measuring previous project’s successes, and the level of collaboration with other freelancers, this dynamic store of data enables new ways to provide quality recommendations.

DREAM Identity is providing the core services for an integrated network of collaboration enabling decentralized teams of freelancers to work together continuously, by automatically taking their past successful deliveries into account when generating recommendations.

Learning from Interactions

Our main goal is to achieve a reduction of work necessary to validate and gather information about the participants on our platform, and to distribute this knowledge back to the community.

This means not only learning about the successes of projects within a given industry or region, but also the best performing team structure, the range and scope associated with a given role and associated professional and experiential background, and even as far as monitoring the impact of contextual influences on these dynamics in near real-time.

All this learning is fed back to the users generating it, and remains under their control with the ability to share it with the community in the effort to learn from each other’s success.

Projects can thereby carry the experience trust of its team and advisors, providing validated information to anyone looking to engage with it.

Experts carry independently validated information about their experience, work-styles, collaboration patterns and skills, and can be automatically associated with challenging and interesting opportunities that go far beyond a simple matching of skill requirements.

Build for growing demand

Within the DREAM Knowledge framework, we provide a series of data streaming pipelines built on the performant structure of our Scala/Spark backend architecture, enabling the use of large quantities of log data for automatic generation of behavioral analytics.

This behavioral feedback loop allows us to not only quantify the current project or freelancer state and recommend a logical next action, or offer a better match, but to learn generalized project success models that are continuously updated.

Beyond a mere statistical description, we aim to highlight critical dependencies between decisions over time, to enable a causal understanding of various actions on a project’s success.

Recent advances in deep-reinforcement learning enable us to model complex interaction sequences in the background, improving on the learned success models from real project interaction, and use optimized policies to enhance the prescriptive aspects of our automated project management engine.

Combined with continuous improvement from our Team of human domain experts, the system can continuously learn to solve two closely related challenges: To assess the likelihood of a projects success given its current state given the relevant context, and a policy to inform the best possible action at this state towards long term success.

Managing Uncertainty

As an entrepreneur provides information to our system, it is naturally incomplete, and decisions are made under uncertainty. This naturally puts a lot of pressure on models that require the full data model to be present as input during inference in production.

Even though regularization and drop out get us a long way to keep models from putting too much importance on individual features, the integration of diverse datasets into a coherent inference stays challenging.

Learning on the Edge

We are equipping DREAM Builder with probabilistic inference methodologies to handle the integration of data sources and the estimation of relevance and impact of a project’s information, all within the given context of that particular stage of the project.

These A.I. systems were previously limited in their scalability by both the complex mathematics required, and the tedious expert dependent work to initialize them. But with recent advances in probabilistic programming languages (which integrate sampling algorithms, deep learning and probabilistic modeling), we now have a new set of tools available to us.

Recent algorithmic advances have been able to significantly reduce the computational expense of approximate sampling-based methods by learning from past inferences on the data. Read this paper for more information.

Deep Probabilistic Programing Languages

knowledge
Frameworks like PyMC3, Edward, and the recently open-sourced PyTorch based Pyro language developed by Uber provide a convenient, programmable approach to generating complex graphical models in a scalable and proven fashion.

These deep PPL's (deep Probabilistic Programing Languages) enable us to start with expert’s models as a pior to initialize the dependency relations, which guide the model in its learning on new data to the areas of highest importance. This way we can continuously update the resulting predictions on real world data we receive from the project’s development, and create generate ground-truth from authoritative third party data sources, which we use to qualify and evaluate our predictions during on-line training and batch training updates to the models.

There is a lot of ongoing research in this area. To get started i recommend you take a look at this tutorial applying PP to financial forecasting, as an example of modeling longitudinal datasources. We will also continue to update on our own research thorugh the DREAM developer series.

Keeping the Expert in the Loop

advisory

As we offer to automate highly individualized aspects of critical planning, fully automated features that are currently beyond the state of the art in even A.I. based chatbots, we apply a layered process architecture to manage and continuously evolve our processes as new solutions for automation become available and our database of successful examples grows.

We deploy augmented Intelligence services in the agent-facing administrative backend, to easily scale the agent's expertise in handling a large quantity of concurrent customer interactions.

The full power of intent classification and entity extraction is made available to the expert agents, which support decision processes in near real-time, continuously increasing the level of automation by learning from each successful interaction.

Smart conversation will “hand-off” to suitable support staff, combined with an integrated scheduling service to match domain expert agents with projects for complex, multi-level intent requests like planning and requirement analysis.

Learning from Expert Interactions

A domain expert fallback is necessary to ensure the best possible outcome for the customer and the model training process, and expert interactions are closely monitored to continuously train and expand the A.I.’s ability for complex, domain specific decision making.

Using a “human-in-the-loop” approach to jumpstart the development of industry-specific guides which are needed to guide the creation of projects enables DREAM to quickly provide validated solutions for our customers.

Our human experts construct logical models which they use to guide a stage-gated process of scoping out the required project work packages from a (typically) loosely-structured description. The work required to complete the project can initially be generalized (and then personalized) as the usage of the platform scales.

Follow our Journey

We are enthusiastically driven to bring the highest level of current technology to team-building and project management, so that the next generation of entrepreneurs can usher in the promises of the technologists that have come before. The future is brighter than ever, and we sincerely hope you will join us in being a part of it.

--Frank Fichtenmüeller, CTO


The DREAM platform isn’t just another untested beta program on a white paper… It’s live and being used right now to hire blockchain professionals. The token sale will enable DREAM’s innovative team to take DREAM to the next level by integrating A.I. and incorporating our platform token.

Social Channels

Part 3 - Using AI to Consistently Deliver Successful Projects and Build Teams
Share this

Subscribe to DREAM Blog