Demand planning analysts at a global semiconductor company worked from forecasts that missed, leaving billions in GPU and CPU revenue exposed. The signals they needed were scattered across systems.
Agentic AI on top of their ML platform, giving analysts a full-stack app that improved forecast accuracy by 40 percent and plans billions in annual revenue with far less manual work.
When a production line stops, operators lose time digging through dashboards and tribal knowledge to find the cause and the fix. Every minute of downtime is lost output.
A generative AI assistant for the floor. An operator asks in plain English and gets the stoppage, the likely cause, and the fix, drawn from structured sensor data and unstructured worker logs. Piloted across three plants of the world’s largest food manufacturer.
An airline’s network planning team knew AI could help but had no prioritized path. Competing ideas, unclear value, and no view of cost made it hard to commit budget.
We ran discovery across the org, mapped the capability gaps, and delivered a sequenced roadmap with a TCO and ROI analysis for each option, so leadership could fund the right work first.
Forecasting when a tropical cyclone will rapidly intensify is one of the hardest problems in weather prediction, and getting it wrong puts lives and coastlines at risk. Operational models had long struggled.
ML.HIFS, one of the first ML frameworks to forecast rapid intensification from satellite data. It beat the National Hurricane Center’s operational model by 30 to 200 percent and earned an AMS award.
An airline had to match the right aircraft to thousands of daily flights under demand, crew, and maintenance constraints. Rigid rules left seats empty on some routes and turned passengers away on others.
An optimization model that assigns the fleet across the schedule to maximize profit while honoring every constraint. It re-solves fast as demand shifts, lifting utilization and protecting revenue.
Supply chain optimization is powerful but locked behind specialists. Planners could not ask their own questions, and moving proprietary data to a model was off the table.
A plain-English layer over the optimizer. A planner asks a question; the system writes the optimization, runs the solver, and explains the trade-off in words, without moving data. A Technical Premier Award asset.
A tennis score tells you who won the last point, not who is in control or whether the momentum has turned. Coaches, meanwhile, lose hours tagging footage by hand to prepare for an opponent.
The models behind the tournament's prediction engine: a win probability that re-scores after every point from 150+ live signals, the performance keys players plan around, and scouting reports built from footage in minutes.
Synchronous CAD promised Google-Docs-style co-editing for mechanical engineering: two engineers shaping one 3D model in real time. But nobody knew how to collaborate well in it, or when editing in parallel helped versus hurt.
A study with MIT and the University of Toronto using a semi-guided BERTopic model over multimodal signals, speech, gaze, facial affect and cursor activity, to map how engineers collaborate and when to switch between shared and parallel modes. Published at ACM CSCW.