Shafik Quoraishee

Shafik Quoraishee is a staff games engineer at The New York Times, where he works on Crosswords, Connections, Strands, Wordle, and Crossplay. A published writer and illustrator with a background in mobile development and computer vision, he investigates how machine learning can make games more interpretable, responsive, and accessible while preserving human-authored puzzles.

Before joining the Times, Quoraishee worked at Insider, where he helped build Produktor, a commerce microservice that separated product recommendations from editorial publishing infrastructure. At the Times, he developed experimental on-device handwriting recognition for the Android crossword app during MakerWeek 2023, combining a custom drawing interface, a convolutional neural network, and TensorFlow Lite. His account of the prototype explored handwriting as an alternative to conventional input, not as an AI feature released in published games.

He also helped build the Connections Reference Dashboard with puzzle editor Wyna Liu. His engineering work on Crossplay’s Android board addressed the practical difficulties of a multiplayer word game: integrating Android Views with Jetpack Compose, moving tiles across a 15-by-15 grid, interpreting imprecise touches, and coordinating drag-and-drop with zoom.

• Human-created puzzles as tests of machine reasoning. Connections requires players to separate 16 words into four groups despite decoys, overlapping associations, and multiple meanings. Quoraishee uses the puzzle to examine whether models genuinely infer relationships, explain their choices, or merely reproduce answers encountered in training. His writing on why people outperform computers at Connections extends that interest in semantic ambiguity and human judgment.

• Graph-based semantic reasoning. His independent Connections research models words as vertices and relationships as weighted edges. Graph coloring reduces the search space; WordNet, ConceptNet, embeddings, and relational-alignment scores distinguish spelling, morphology, cultural associations, and polysemy. A graph convolutional network proposes groupings, with reinforcement learning exploring candidate solutions. The work is preliminary and separate from internal Times research.

• On-device game agents with measurable constraints. In joint work with Joanne Song, Quoraishee developed experimental Space Invaders and crossword-solving agents while examining local inference, offline availability, privacy, and personalization. He focused on how an agent can share memory, battery, and rendering capacity with a mobile game, including the approximately 16-millisecond frame budget of a 60-hertz display. Song contributed the presentation’s accessibility framework; the prototypes were not deployed Times game features.

Quoraishee’s independent projects also include BLSNet, which explores labor-market data, and Foodlifier, an image-based food-recognition experiment.