Log

The first machine was a computer running code I did not fully understand. I started with machine learning the way most people do — from the top down, fitting models to tutorials. It worked, in the sense that loss went down. It did not explain anything.

The turn came when I started asking what was underneath: what a gradient actually was, what the hardware was actually doing while the notebook ran. That question — what is happening underneath the abstraction — became the organizing one.

I built small things badly and rebuilt them. The phase produced no portfolio worth showing, but it produced the habit everything since has run on: take the opaque thing apart and rebuild it from first principles.

In 2023 I stopped following tutorials and started running experiments. The repos from that year are honest about their scope: Face-Regeneration, Detect-color-signs, Track_and_Predict, a decision-tree classifier — small projects whose real subject was learning to use a computer to see stuff.

The comma.ai calibration challenge was the first time I worked against a real stack. Calibration is the unglamorous core of any camera system — estimating the intrinsics and extrinsics that turn pixels into geometry — and the challenge made me sit inside it. There is no magic, just long calibration.

The Duburi interview task was the door into the team. I built it, and it opened the next phase.

Joining Duburi’s AI and machine-vision subteam meant my code stopped being an exercise and became load-bearing on a vehicle that has to perform in water that does not care how elegant the code looks.

The subteam taught me that perception is a pipeline with budgets, not a model with an accuracy number. Frames have ages. Buffers have latency. The detector can be perfect and the vehicle can still fail, because the information arrived too late to steer.

I owned pieces of the vision stack. When the robot misbehaved, the cause was rarely where the symptom pointed. That lesson — measure the actual path, not the assumed one — is what I carried out of the phase.

The step from member to lead changed the work’s shape. As AI and machine-vision subteam lead, I owned a pipeline other people depended on; as engineering lead, I owned the vehicle as a whole — a different discipline entirely.

Leading meant the failures became public. A system that works on the bench and fails in the water is not a system that works. The 2025 RoboSub competition ended 9th — a number that lives as lessons about integration, timing, and the gap between components that work and a vehicle that works.

The phase taught me that engineering leadership is mostly making dependencies visible: who needs what from whom, and what happens when it is late.

The work moved from the vehicle to the system around it. duburi_ws became the ROS2 control, mission, and simulation stack; duburi-sim_ws became the Gazebo + ArduSub simulator where reality is approximated before it isn’t; duburi-codebase_RND became where experimental ideas get tested on Duburi 4.2 before earning a place in the stack.

The 2026 RoboSub season ended 8th with 6230 points — a step up, and evidence that the systems work was compounding.

The university detour made the path non-linear, but the engineering became increasingly serious. The phase’s lesson: a vehicle is only as good as the tooling around it, and the simulator is where you learn what you do not know about the water.