Forecasting, machine learning and AI agents that hold up in production#
I help data, science and technology leaders turn a model into a system their organization relies on: independent assessment when you need a second opinion, hands-on delivery when you need it built. I'm Michele Dallachiesa, Ph.D., an applied scientist based in Munich, Germany and independent since 2019. Before that: Lead Data Scientist, founder and CTO.
Clients include
How I can help#
Three ways to engage, matched to where your team is. Most engagements start with an assessment.
-
Assess
An independent, evidence-based answer before you commit budget. I review methods, models and vendors, measure what an existing system actually does, and design the evaluation before the build.
Examples: a review of a contact tracing system whose recommendations were implemented, and a feasibility study on AI for legacy codebases whose findings shaped a venture's product direction.
-
Build
Forecasting models, recommenders, LLM evaluations and agent systems taken from problem definition to production, inside your team and on your stack, with a report at every milestone.
Examples: cost forecasts for a railway megaproject used by senior leadership, a recommender serving millions of researchers, and contact center forecasts that cut error by a third in production.
-
Enable
Your team owns the result. Every build ends with reproducible code, documentation and a handover session, and workshops bring sales, management and engineering up to speed on machine learning.
Examples: a workshop for the sales and management team of a mobility provider, and the upskilling of a publisher's engineers on the recommender systems they now run themselves.
Selected results#
−⅓
lower forecast error on weekly phone contact volume, in production with weekly retraining, with estimated savings in the millions per year.
−20%
lower refund-estimate error for a tax-filing app, with calibrated quantile predictions deciding which payouts go to manual review.
−10%+
lower error on cost and schedule overrun forecasts across hundreds of public capital projects, with conformal uncertainty bounds.
I wholeheartedly endorse Michele and his efforts, as his contributions have been instrumental in the successful implementation and deployment of the forecast models.
What I work on#
-
Agents that triage security alerts, answer questions over warehouse and HR data, and assemble research datasets from public sources one sourced cell at a time, each measured against ground truth by an evaluation harness.
-
Recommenders built on fine-tuned LLMs and embeddings, text classification for regulators, multilingual speech-to-text evaluation, reusable retrieval-augmented generation components, and verifiable AI inference.
-
Cost, schedule and demand forecasts for megaprojects, water utilities, grid operators, contact centers and manufacturing, with conformal prediction and explainable risk rankings that tell you where to look first.
-
Mobility and location analytics
Traffic simulation and signal optimization, public-transport planning from fare-card data, an independent review of a contact tracing system, and GDPR-compliant movement analytics at billions of records a day.
How an engagement runs#
1. Discovery call A short conversation about the problem, the data you have and the decision the result must support. You get a frank view on whether it is worth doing.
2. Written scope A proposal with deliverables, milestones and evaluation criteria agreed upfront, so everyone knows what "working" means before any of the work starts.
3. Delivery in your team I work on your data and your stack, report in writing at each milestone, and flag what the data cannot support as soon as it shows up, not at the end.
4. Handover Reproducible code, documentation and a working session with your engineers, so the result keeps running and improving long after I have left.
I work remotely from Munich with clients across Europe, the UK, the US and Asia-Pacific. Engagements range from a few weeks of assessment to multi-month builds, and several clients have come back for a second phase.
Background#
-
Research
Ph.D. in Computer Science from the University of Trento, research stays at the IBM T.J. Watson Research Center and the Qatar Computing Research Institute, and 11 publications in VLDB, SIGMOD and other venues.
-
Built and led
Lead Data Scientist at Minodes, a Telefónica NEXT company, and founder and CTO of Skysense, which pioneered charging stations for commercial drones and sold to a space agency, government agencies and universities.
-
Speaker
Regular speaker at international events, including PyData Berlin, MLOps Munich, ML Week Europe and EuroPython, on forecasting with calibrated uncertainty, evaluation, and machine learning systems that survive production.
-
Full stack
From model design and LLM evaluation to Spark, Snowflake, Kafka and FastAPI on AWS and GCP, so a model ships with the pipeline, API and monitoring it needs to stay in service. See the full technology stack.
The full list of projects is on the Experience page and my publications are on the Research page. For my career before 2019, request my two-page CV.
Get in touch
The quickest way to find out whether I can help is a short call. Prefer email? Write to [email protected].
Book a discovery call Connect on LinkedIn Subscribe to the newsletter