Stefan Atev Email & Phone Number

Remote (Hq In Maple Plain, Mn)

Responsible for toolpathing, probing, and CAD application integration of Proto Labs' in-house CNC Mill/Turn CAM package. Created DSLs for type-safe G-Code output, algorithms for 3D collision detection and spatial queries, as well as feature detection for feature-based toolpathing. Help refactor large swaths of the C++ application and libraries in order to unify CNC Mill and CNC Lathe support.Created prototype ML algorithms for instant evaluation of CAD models for cost estimation / manufacturability metrics. Apart from generic geometric features, used fast machining simulation algorithms to inject CNC machinability constraints into the developed models. A variant of the machining simulation I developed is also used for toolpath verification and optimization in internal applications.I developed several enhancements to Proto Labs' MIM (metal injection molding) and Thixomolding automatic quoting process - algorithms for estimating the quantity of MIM parts that fit in a sinter tray, and for estimating the maximum material flow length during magnesium casting. Implemented high-performance interactive collision detection used while designing machining fixtures for molded parts.My first experience in toolpath algorithms was to improve automatic tool selection and to vary tool feed rates based on cut types. Developed interactive visualization for uncut material during automatic and manual tool selection.

Minnetonka, Mn

In a slight departure of my scientific duties, I led the redesign and implementation of Vital's data loading and normalization pipeline, which aims to standardize and validate DICOM metadata from various vendors. My focus in the redesign was to use compile-time type-safety in client APIs and to explicitly audit metadata changes in order to minimize the potential for incorrect/unrecognized modifications. Developed a very deep understanding of the DICOM standard as a result.I developed a high-performance algorithm for non-rigid 3D registration that is diffeomorphic and inverse-consistent, with hyperelastic regularization. The algorithm uses second-order optimization techniques to improve convergence and has been especially successfull in applications where high-resolution registration is necessary (i.e. digital subtraction angoigraphy, etc)To address challenges in low-dose imaging, I developed an edge-preserving noise reduction algorithm that is fast enough to allow interactive control over the denoising parameters, giving physicians finer control over the results than competing methods. An extension to the algorithm that takes into account image structure other than edges has proven particularly well-suited for CT lung, cardiac, and surgical imaging.

Minnetonka, Mn

My experience at Vital started with an urgent assignment to optimize the 4D brain perfusion application - it had to process about 4GB of dynamic images in under 2 minutes. I managed to bring down the processing time from over 3.5 minutes to 45 seconds (on same hardware) and in the process became intimately familiar with deconvolution based and model-based methods for perfusion computations. I have traveled on several occasions to educate neuroradiologists and neurosurgeons on the specifics of Vital's delay-corrected perfusion algorithm.I developed a novel parallel iso-surface extraction algorithm that is lock-free and fully vectorized. It is used in Vital's generic region editing and can produce and smooth a 10M triangle mesh with correct topology in about a second.Binary morphology is a workhorse of segmentation methods. I co-developed an optimized morphology algorithm for arbitrary structuring elements that heavily uses SWAR (SIMD Within A Register) techniques to achieve high performance.All projects have required expert use of OpenMP and SSE intrinsics, as well as a lot of careful profiling with Intel VTune.

Minneapolis, Mn

I worked on a variety of projects at the Artificial Intelligence, Vision and Robotics Laboratory. The main focus of my work was on developing video analysis software for both transportation-related research (funded by Mn/DOT) and for threat assessment software (funded by the DHS and related agencies). Unlike typical university projects, both of these efforts had deployment of production-quality software systems as deliverables.During the course of my work at the lab, I have used and implemented algorithms for motion segmentation (using Gaussian mixtures and MRFs), dense optical flow, multiple target tracking (appearance and region-based methods), on-line PCA (linear and kernel), spectral clustering, SVMs, Bayesian networks, Discriminant Analysis (linear and kernel), intrinsic image recovery (shading and reflectance decompositions), illumination-invariant image filtering, object detection. While a lot of these methods were initially prototyped in MATLAB, I have implemented all of them in C++ as well.

Minneapolis, Mn

I led the weekly recitations (2 lectures of 50 minutes each) for CSci 2011 "Numerical Methods", the introductory class on numerical analysis in the Computer Science department. I made it a point to visit every regular lecture for the class together with the students.I received very good reviews from the students who specifically liked the efforts I made in explaining everything in different ways and the fact that I focused on areas I had noticed were cause for confusion during the regular lectures.

Cambridge, Ma And Decorah, Ia

Worked remotely (part-time) during my junior year and first semester of senior year of college, and full time during the summer of 2002 and from January 2003 to July 2003. Initially worked on SashXB (see Extreme Blue Internship). My biggest project was related to Semantic Web technology with the RDF-based Life Sciences Identifier (LSID) project, which is now an OMG standard. I implemented parts of the reference LSID client and server implementations. The people I worked with have since moved on to Cambridge Semantics.Used a lot of Java and Perl to design web services. Showed a prototype graphical browser for RDF I developed with Java and SWT at a booth at BioIT 2003.

Cambridge, Ma

My first experience at Big Blue. Worked on SashXB (Sash for Linux) in the Cambridge, MA campus (Lotus Building). Told Sam Palmisano during his visit to Extreme Blue, and I quote from memory, "Linux is the best thing since sliced bread". We were using Mozilla's JavaScript engine and did three full rewrites since Mozilla had not hit 1.0 yet and were changing their interfaces every point release.I used C++ a lot and got familiar with XPCOM, Mozilla's cross-platform version of COM. Since XML was all the rage, I wrote an XPCOM component for easy XML parsing (version 1 used Xerces, version 2 used Gdome). Learned the ins and outs of GTK+ and Gnome development as our project was open source and hosted on Gnome CVS.

Ph.D., Computer Science, 4.0

Activities and Societies: Best paper award for "Kernel Spectral Curvature Clustering (KSCC)" at ICCV Workshop on Dynamical Vision.

M.S., Computer Science, 4.0

Activities and Societies: ACM Intercollegiate Programming Contest - World Finals 2004Thesis project: Real-time vehicle tracking, collision.

B.A., Computer Science, Mathematics, Physics (Minor Field), 3.98, Summa Cum Laude

Activities and Societies: Phi Beta Kappa, Sigma Pi Sigma, Pi Mu Epsilon; Meritorious, COMAP Mathematical Contest in Modelling, 2002; ACM.

Mathematics And Computer Science

Activities and Societies: Several 1st place finishes in national programming contests and Olympiads.