Luca Fiaschi uses Bayesian models when forecasts need interpretability, calibrated confidence intervals, and constraints such as positive outputs.
2
A LangGraph application can connect quality-control, modeling, forecasting, scenario-planning, reporting, and presentation agents, with a business-context expert checking the results.
3
Synthetic consumers can test product concepts and user experiences, although prompting works better for broad populations than for fine-grained user segments.
Summary
Luca Fiaschi describes how he combines traditional machine learning, Bayesian modeling, and generative AI to speed up analytics work. Bayesian models help explain forecasts, express uncertainty, and add constraints in small-data, high-stakes settings such as marketing allocation. His team is building a LangGraph workflow in which agents check data, produce insights, build models, run forecasts, plan scenarios, and prepare reports. A human still needs to judge whether the data and outputs make business sense. Fiaschi also discusses synthetic consumers, which are language-model-based profiles that can react to product concepts, prices, images, websites, and user flows. Prompting can capture broad population behavior, while supervised fine-tuning and model adjustments may help with narrower segments. He is candid that evaluation remains unsettled. The conversation ends with his approach to finding business value: understand the business model, identify variables with high elasticity, and give analytics teams explicit opportunity-focused goals.
Bayesian models make high-stakes forecasts easier to explain
Fiaschi says he moved toward Bayesian models because stakeholders wanted to understand the logic behind forecasts, the prediction interval, and the variables involved. Traditional machine-learning models can provide explanations, but he finds them intrinsically hard to explain and their confidence intervals are sometimes poorly calibrated. Bayesian models provide interpretability and confidence intervals more directly. They also make it easy to add constraints. If an output must be positive, that can be built into the model. He gives marketing allocation as an example, where HelloFresh spent $800 million across 40 or 50 media channels when he worked there.
Agents can extend analytics workflows from model building to business questions
Fiaschi describes using LLMs to augment data scientists and analysts. Some agents act as modelers, connecting variables and writing code. Others check incoming data. After a model is built, users can ask for forecasts or what-if analysis in ordinary language, such as what would happen if the marketing budget for Google dropped. The model can answer without an analyst manually rerunning code. Fiaschi frames this as a way to address the difficulty of scaling data teams and delivering follow-up analysis to stakeholders.
The workflow is a graph of specialized agents with a human business reviewer
Fiaschi says the current system is built as a LangGraph application. A quality-control agent checks the data, an insights agent creates plots and describes trends, a modeling agent builds the model, and forecasting and scenario-planning agents use it for future estimates and optimized configurations. The team is also building agents that create PDFs, text, and PowerPoint decks for stakeholders. Automated checks can find missing values, unusual scales, or implausible inputs, such as $200 million spent in one marketing channel in a week. A human with business context still has to judge whether the data and final output are sensible.
Synthetic consumers can test products before real consumer research
Fiaschi describes synthetic consumers as virtual representations of customers that people can question about products and user experiences. A consumer packaged goods company could show a synthetic panel a toothpaste concept and ask whether people would buy it at different prices or which features they like. A product team could send an agent through a website, ask whether a button was confusing, and learn about problems in a user flow before bringing in real users. Fiaschi says research suggests these systems can represent real populations within certain constraints, provided they are built with suitable data.
Prompting works for broad groups, while narrow segments need more work
The team is experimenting with prompts that specify demographic and behavioral details, such as a Black woman living in Brooklyn with a tech job and recent purchases in a product category. Fiaschi says prompting gets the system roughly 60 to 70 percent of the way there. It captures the average and bulk of an American population more effectively than smaller subsegments. Historical consumer research and surveys can support supervised fine-tuning, using open-source models such as Llama. He also mentions investigating ablation techniques to reduce model biases and make the base model more impartial.
Agent evaluation has reference workflows, but no settled industry method
Fiaschi says his team is still learning how to evaluate these applications. They use reference data and workflows, then check whether agents reach expected conclusions, avoid getting stuck, and produce model parameters aligned with reference parameters. Synthetic data can help verify some of these cases. LangSmith and other telemetry provide information about user queries and where users get stuck. He does not claim the problem is solved. Finding a systematic evaluation method for this class of applications remains an open question for his team and the wider industry.
Analytics teams should search for opportunities tied to business elasticity
To connect AI work to business results, Fiaschi first studies the business model as a graph of variables. He looks for variables where a small percentage change produces a large change in revenue or profitability. He also treats analytics teams as opportunity finders because they work closely with stakeholders and can notice needs that platform and machine-learning teams may miss. One KPI he gives analytics teams is to identify $10 million or $20 million of new opportunities by the end of a quarter. The exact amount matters less than putting the team on the offensive to find and develop opportunities.
The right AI projects depend on the company's growth stage
Fiaschi says a fast-growing company should usually prioritize growth problems over small automation or cost-saving projects. Spending a year to save four people ten hours each may have a poor opportunity cost when the business is expanding quickly. At a company where growth is difficult, cost reduction can matter much more. He gives a large call-center operation as an example where automating customer-success work could affect a substantial cost base. Potential use cases need to be judged against the company's current strategy and stage.
"It needs to be a human that has business context good enough so that can understand whether the data is the right data and the output of the entire workflow is something that's sensible."Luca Fiaschi17:09
Who should watch
You run an analytics or data science team and want agents to handle model setup, data checks, forecasts, and follow-up questions.
Product teams are considering synthetic research and need a grounded account of where synthetic consumers may match real users and where they can fail.
You need to choose AI projects by business value and want a practical way to connect analytics work with revenue, profit, growth, or cost.
