Work
Ten case studies.
Each one is a sequence: the problem, the design, the implementation, and the benefit.
Pending home sales forecast
Operators needed a national pending-home-sales figure before the actual count was published. The series turns with the seasons, so a flat guess is wrong even when the level looks close.
CAM recovery likelihood
Common-area charges are not recovered evenly across locations. A score fit on the same year it is judged flatters the result.
Gold futures backtest
A futures rule can look fine in total and still depend on one or two years. The review has to show the result and the number of trades, year by year.
Trade-indicator orchestration
Indicator readings were one step away from an action. Nothing required a risk check to pass, and nothing required a person to send.
Trading algo decisioning
A trading algorithm needs a draft and a classification. If the same component can also send the order, a stale snapshot or a repeat can act twice.
Real estate development platform
Listing and transaction feeds arrived continuously and were easy to edit in place. Pending and closed records were easy to mix. A forecast could reach consumers even when it failed its check.
Equity-index futures research lab
The question was an opening-range break on those four contracts. The lab had to study that event without a path to an order.
News ingestion, clustering, and summaries
The same story arrived as many articles. A summary per article repeated the story and mixed in sources that should have stayed out.
Cultural-model evaluation - open-weights fine-tuning
A public model can be adapted to a cultural task. Without a frozen copy of the same starting weights, there is nothing honest to compare the fine-tune with.
Bet-log ingestion and sequence model
Bets lived in a manual log. There was no ordered history per account, and no separate step that scored the next outcome.