Abhikalp Unakal

Original Research

Katana, utility AI

[Code] · [Paper]

A 2D football simulation built in C# and Unity using utility-based behavior trees, simply because of my love for the game. This was original research work I did while at DigiPen, where I spent the better part of a semester studying football game dynamics, physics, and game tactics from first principles and figured out how to mathematically model football so that I could give the computer actuator knobs to attack, defend, dribble, and kick. These behaviors come from custom hand-tuned objective functions that can be dynamically modified, rather than a hard-coded football AI. I conducted independent research simulating, from physics first principles, what happens in a game of football. I reached out to several professors across the world with questions when I’d get stuck, and finally I built a product and tested it against some of my FIFA maniacs, who said the AI was quite impressive for a one-man project and in some situations better than the original game AI. There’s also a YouTube video of me explaining what I built above.

Video Games

Space Jump, 2D puzzle platformer powered by original game engine Atom

[Code]

Video of Atom game engine.

Screenshots from the actual game.

I was director and producer of Space Jump, a 15-level spatial puzzle with a teleport mechanic, powerups, and boss fights. I was also tech lead on Atom, the C++ engine it runs on, with three other engineers. The engine featured 3D spatial audio, an event system, and a highly efficient and memory-dense ECS. Super proud of this project, as it was the first game team that I led. Not only did we ship an entire game from scratch, but I also architected and built an entire custom game engine from scratch. It was the first time I conducted user playtesting and observed the unique ways in which people would play the levels or get stuck at harder puzzles, which provided us with useful signals on how to modify the difficulty curve for the level progression.

Work

Audio research

• Shipped [Conversation Focus] on Ray-Ban Meta glasses.
• Architected metrics and telemetry for the glasses and research prototypes. Live tuning of DSP parameters. Moved the team from monthly checks to continuous hourly validation. C++, C#, Python.
• Led a petabyte-scale multimodal collection warehouse: custom networking, hardware, several sensor types, motion capture, metrics, and telemetry into deep-learning annotation pipelines.
• Built a C++/ImGui audio-graph visualizer that generates DSP pipeline diagrams, like Blueprints for an embedded signal chain, and flags breaks before they ship.
• Upgraded unidash from 1D to N-dimensional plotting in Python and Hack, so frequency-response curves became a first-class view for about 80,000 engineers.
• Built the packaging and licensing layer that let Matlab research code run in the cloud, so research and product could check feature parity.
• Ran a research experiment on the capture rig: a C# and Unity hardware-validation system that tracks head pose, azimuth and elevation, with a fine-tuned YOLOv3 model. It aligns video frames for photogrammetry and removes the need for a research assistant to position the participant by hand.

• Led UX engineering on the Roland Cloud editor team. Designed the interaction and built a programmatic auto-layout engine for the [JUPITER-X editor], so players can manage sound profiles.

• I worked as a platform engineer for the advanced analytics and platform team at Danske Bank, building deployment pipelines for credit-risk machine learning models in Spark, and a custom heartbeat layer plus CI/CD in Jenkins on a Hadoop cluster and Linux boxes, so regressions were caught before the bank’s customer-facing teams flagged them.