I figured I'd give occasional blogging a go, and this will be my first post. While I don't have a detailed plan for this space, I intend to share professional updates that are top of mind for me here, such as career updates, blurbs about new papers (esp. my own papers!), and possibly general musings about where things seem to be headed with market design, supply chain optimization, AI, RL, and operations. This post will be a brief one about what I managed to get done so far at Amazon, where I spent the last year doing supply chain optimization full time. The sabbatical gave me the chance to both learn and make a substantial impact.
At Amazon, I led the creation of a general-purpose primal dual algorithm for solving very large resource allocation problems (millions of decisions, with hundreds of thousands of constrained resources), e.g., for assigning customer shipments to "paths" in the supply chain. The new general solver works an order of magnitude better and faster than the previously used custom algorithms, and is already being used by two key supply chain decision systems, with more in the pipeline. This work won the Operations Research Best paper award at Amazon's Consumer Science Summit 2023 (1/108 submissions). A primal dual algorithm, informally speaking, solves a resource allocation problem via a "negotiation" between millions of orders/shipments on how to share scarce resources (e.g., transportation capacities, sortation capacities, last mile capacities). My brilliant former PhD student, Pengyu Qian, taught me primal dual algorithms. I leveraged his mirror backpressure concept in my algorithm design:Y. Kanoria and Pengyu Qian, “Blind Dynamic Resource Allocation in Closed Networks via Mirror Backpressure,” appeared in ACM EC 2020, to appear in Management Science.
A shout out also to my colleague Santiago Balseiro, who has successfully used similar ideas in a different application, and shared some of his experience with me.
A shout out also to my colleague Santiago Balseiro, who has successfully used similar ideas in a different application, and shared some of his experience with me.
One other big thing seems to have landed at Amazon. I discovered a major inefficiency in how customer orders were being assigned to fulfillment centers. The new decision framework I proposed will be simpler, more transparent, and efficient, and is expected to save nearly half a billion dollars/year. My proposed system will make tradeoffs consistently, allowing to operate at the system-level Pareto frontier.
I'll be continuing my work at Amazon part time.. there's billions of dollars (and proportional amounts of carbon) still to be saved, and lots more for me to learn! The two projects I described above were perhaps among the lowest hanging fruit that I identified, and I was able to move those forward in less than a year after I formulated the underlying idea. I have made a number of additional proposals which are being considered (or not, depending on the weather that day) which are more ambitious but also more complex to assess and roll out. Having achieved some success, I sense more trust and backing from leadership, and I'm hopeful that some of them will move forward in time. If I get a chance, I'll share my experience this past year with organizational inertia and the trust building process in the industry in a different post. Overall, I'd say that identifying and "solving" scientific problems has been the quick and easy part of my journey at Amazon, which is obviously very different from my experience as an academic (though peer reviewers can certainly be a tough bunch!).
I'm super excited to be back full time at Columbia (since Sep 1), with my students and colleagues, focused on research and teaching. The AI wave seems to be rapidly engulfing all things in my professional sphere, and I'm curious to see where it will go, and how I can contribute.
