Past Consulting Projects
Predictive Models for Electronic Devices. EmpiricalLinear Prediction is a technique for reducing testing costsfor devices such as digital/analog converters by replacing alarge fraction of the measurements with predictions andrigorous uncertainty bounds that are based on computationalmodels. The models are obtained by applying datareduction techniques to data from suitable training sets. Acomprehensive method for this purpose, HighdimensionalEmpirical Linear Prediction (HELP) was developed in thelast decades at NIST. Ideveloped, tested and implemented numerical algorithms,statistical models, and graphical tools for the HELP toolbox.This work was done with a team of engineers and statisticians atNIST and academic researchers at Cornell and NortheasternUniversity.
Clustering Approaches for IDDQ Data. IDDQ testingis a modern test technique for electronic devices that is basedon measuring the quiescent supply current. These data can begathered very rapidly and can be used to complement andreplace other more expensive testing methods. In thisproject, carried out at NIST with a team of engineers, Iapplied data mining techniques such as clustering, principalcomponent analysis and HELP (see above), and logistic regressionto a dataset obtained from the SEMATECH consortium, with the goalof predicting the outcome of conventional tests from IDDQdata.
Frequency Responses from Step Response Data. Thefrequency response of an electronic device can in theory bedetermined from its response to a voltage step, if it measured atpicosecond intervals. While such measurements are now technicallypossible, practical difficulties arise from the need to handleenormous data sets (many millions of test points if the devicehas to be characterized from the kilohertz to the gigahertzrange), and also from measurement uncertainties which areamplified during the computational process. I developed awavelet-based noise reduction scheme together with afrequency extraction method and a method for obtaininguncertainty bounds for non-uniformly sampled data thatrequire fewer measurements. As part of the project, I identifiedthe underlying mathematical principles and tested and implementeda suitable set of computational algorithms. The work was done atNIST.
Work as Statistics Expert Witness. In a case before theUS District Court for the District of Columbia, I was retained bythe Defendant's counsel to evaluate statistical evidencepresented by Plaintiff's counsel. I evaluated several designs forobservational studies and carried out computationalevaluations of several large datasets (20,000 and morerecords), using spreadsheet and statistical software. I wroteseveral reports and was deposed twice.
Evaluation of a Proposed Public Key Cryptography System.I evaluated a newly proposed public key cryptography system,determining its theoretical and computational feasibility and itssecurity from attacks and writing Mathematica routines to attemptan attack.
Patterns and Trends in Customer Satisfaction Data. Results from monthly customer satisfaction surveys for a manufacturer of medical devices showed some unexplained trends. I analyzed these data with methods from statistics and data mining and identified early indicators of these trends. Jointly with Kim Sellers.
Analysis of shipping data for manufacturing association to prepare for shipping tariff negotiations. I analyzed shipment data for the National Paint and Coatings Association prior to negotiations on truck tariffs.