How do you optimize speed and quality in AI project delivery?

To accelerate delivery you should help your stakeholders to know the basics of AI. Educate them on the model you choose, how it can be tuned and what limitations it has, at a high level, this should be enough to secure you from changes in requirements, demos that did not meet expectations, as well as delayed or canceled projects.

In my experience limiting scope provides foundations for speed and quality. Focussing on essentials allows for a minimal feedback loop. As a result of this iterative approach, we may be able to try different approaches, improve our models, and increase their efficiency.

To accelerate the delivery and improve the quality of AI projects, start by outlining crystal-clear objectives, including expected outcomes, stakeholders, and success metrics. Choose a flexible methodology like Agile or Scrum that encourages quick adjustments and iterative progress. Data quality is non-negotiable; invest in data cleaning and validation to avoid skewed AI results. Please monitor key performance indicators to spot errors or biases in real time, making immediate adjustments to your AI model as needed. Employ AI-specific performance tools for ongoing optimization. These steps create a streamlined, quality-assured process for delivering high-impact AI solutions.

We need clear objectives as well as predefined actions for predictions. Consider an AI model predicting pregnancy blood loss risk. What actions would be taken based on predictions? Possible actions: Referring women to facilities with blood banks Sending the patient's blood group type to her facility. However, models produce false positives & negatives. With referral actions and high false positives, blood bank facilities could become overburdened, impacting all patients. Sending blood allows flexibility for false positives while reducing false negatives. Knowing the actions helps data scientists build impactful solutions. The goal of an AI project should extend beyond making good predictions. It should focus on the quality of outcomes.

Just like any other project, clearly defining and agreeing on the definition of done is crucial for an AI project too. 'Starting with the end in mind' is a gospel. A charter or project kickoff document should outline topics in scope, more importantly topics which are out of scope, outcomes, deliverables, define the high level implementation strategy, identify, list and draft the RACI matrix, define project milestones, list the communication plan, and highlight known and foreseeable risks.

Data Availability and Data Quality are the key foundation for output quality and reliability of AI models. Working on having clean and high fidelity datasets and scoping this work into project delivery timeline/scrum can help optimize speed and quality in AI project delivery

Developing AI models differs from traditional software development. I advocate for flexibility to embrace experimentation through shorter sprint cycles, to allow for more frequent changes. Also, a clear Definition of Done, to avoid open-ended experimentation. Dedicated one-off "innovation sprints" may not provide enough experimentation for AI teams. Management styles should factor in the team's role: Is the data science team part of a single AI product? Or are they a standalone group supporting multiple products across various value streams? The nature and goals of the team should inform the management framework, to balance creative learning with delivery responsibilities.

It’s very important to have a flexible approach to methodology selection when approach any AI project implementation. Consider not only the style of project management to be undertaken, but also the various organisational levers that need to be managed around culture, security, compliance etc. Given the sensitive nature of data, and the expansive nature of AI to touch many organisational points simultaneously, the earlier you consider this in the project life-cycle, the better.

Gathering an understanding of the state of underlying data, the models which will be used, play a crucial role in determining the methodology and tools needed. The initial few days will need to be spent understanding data quality, integrity to determine cleansing processes and ascertain data availability to define next steps and tools needed. In parallel, running POCs to test out hypothesis and determining testing strategies all play an important role in determining tools and methodologies needed.

I've found that communicating such early and often (regarding methodology and tools) sets a project up for better success. Ensuring each sprint has both follow through and the ability to influence the next sprint without lag is key. Informing stakeholders of expectations limits unforeseen challenges to the process, especially when done early and often.

Monitoring data and models is crucial in AI projects. For data, keep an eye on data drift, which refers to changes in data distribution over time. Regularly audit data quality to catch missing values or outliers and ensure relevance. Track data lineage for accountability and ethical compliance. For models, monitor performance metrics like accuracy and precision to detect model drift. Anomalies in predictions could indicate issues requiring retraining or adjustments. Use versioning for data and models to keep track of changes, aiding in troubleshooting and ensuring transparency. These practices ensure your AI project remains robust and up-to-date.

Keep an eye on your cloud or hardware expenses. This might be a great signal to you that you are managing your data wrong and might need to use another model.

The following stages should help a good culture of innovation: 1. Consider the existing culture and its roots. 2. Canvas the team and stakeholders to understand their priorities. 3. Engage in open communication to communicate goals, while also listening to concerns. 4. Be transparent and always communicate the underlying purpose. 5. Highlight opportunities for shared value, mutual goal alignment and make space for concerns and input. 6. Ensure a safe and open climate that gives all members the security they need to be themselves and really engage with the effort. 7. Reward innovators and exemplars on the team by giving them the applause they deserve. Great leaders give credit when and where it is due.

AI projects differ from traditional projects, it requires a lot of experimenting, PoCs, learning cycles and curves. Hence such projects cannot be tied down by very stringent timelines and there has to be scope for failure with open arms. It is when the team isn't scared of failing,.when they'll thrive and innovate, and is something that projects stakeholders must acknowledge abd agree. Rewarding and motivating the team is very crucial.

I agree with this statement; in almost all areas of life, embracing the learning opportunities available in what feels like a failure is sometimes where the true solution lies. Often times, the true depth of needs and expectations are found at the end of one failed project, or more optimistically-the start of a new opportunity.

Consider AI Governance and AI Regulation challenges early on. By understanding the concerns senior stakeholders may have, and how the initiative needs to fit in the organisations AI governance structure, you are better placed to make the AI initiative successful from the get go.