- Co-creator of IDrive Autonomy: led full-stack development and deployment of a mapless L2+ ADAS on a drive-by-wire Ford Mustang Mach-E (NVIDIA Jetson AGX Orin, ROS 2), validated over 1,128+ mi / 25+ h of engaged public-road driving up to 65 mph.
- Own the real-time C++17 planning & control stack (~14.5k LOC, 50 Hz): acados nonlinear MPC path tracker at 72.9 µs/solve, curvature tracking RMSE 3.5 × 10-4 1/m (r=0.97), and an online steer-offset estimator within 0.04° of GNSS/INS ground truth.
- Designed the longitudinal driving behavior: IDM car-following with time-headway gaps, a stop-and-go / crawl / launch state machine, cut-in handling and lead arbitration; 12× fewer out-of-envelope commands than a PD baseline.
- Built an anticipatory curve-speed governor with hold-and-release hysteresis, eliminating speed hunting on road; across 142 replayed curves, p95 lateral acceleration fell 1.56→1.47 m/s2 and post-apex speed loss 1.34→0.85 m/s.
- Found from road logs that the velocity interface braked by regen only (~1.5 m/s2); designed a supervised direct-brake arbitration layer (up to 5 m/s2, seamless hand-back): in 306 safety-critical closed-loop replays, collisions fell 157→0 and median min-TTC rose 0.22→3.7 s.
- Built a 4D-radar multi-object tracker (Kalman with Doppler updates, 18 Hz): ego-speed estimate r=0.98 vs. CAN, stationary-lead speed error 1.10→0.23 m/s, and vision-guided lane assignment lifting correct far-lead selection in curves 54%→87%.
- Developed a zero-phantom stopped-vehicle detector (elevation gating, multipath-ghost suppression) confirming in-lane stopped vehicles at ~100 m vs. ~40 m for fused perception, with 0 false detections over 1000 miles of replayed drives.
- Built the data-driven development loop: per-drive scorecards, actuator system identification from road logs, closed-loop log replay gating every planner change, crash-safe logging, and one-command deploy with self-test and rollback.
- Ran weekly supervised road tests and drive reviews, root-causing every takeover; drove 11 evidence-backed requirements to the perception team (2 shipped, e.g. radar-range retention for braking leads 65-75%→99.6%).
Building the self-driving stack - perception to actuation.
I'm DeeKay Goswami, an AI Engineer (Autonomous Vehicles) and co-creator of IDrive Autonomy - an in-house scalable L2+ codebase. My work starts downstream of vision model inference (a 10.0 s trajectory horizon): I fuse classified vision detections, then turn them into real-time motion - my RADAR object tracking, planning, and C++ vehicle control.
Cruising
◀▶
Simulation of real on-road performance - lane keeping, adaptive cruise & signalled lane changes
55mph
SET55mph
- Gap-
- Headway-
- In-lane0.0 cm
- Curvature+0.0000 1/m
Ego
Lead
Traffic
Planned path
72.9 µsacados nonlinear MPC path tracker, per solve
0.98correlation of the 4D-radar tracker's ego-speed estimate with CAN
1,128+ miengaged public-road driving (25+ h, up to 65 mph)
01
About
I turn perception into motion. The vision model hands me a trajectory over a 10.0 s horizon; alongside it I fuse detections into object tracks, plan through them, and hand the result to the controller. Upstream, I built the training-side planning (a hybrid A* planner over a local costmap) that generates the vision model's labels; downstream, its output drives the on-car planner and control, so my work spans the full stack from label to actuation. At IDrive Inc. I co-created a mapless L2+ ADAS on a Ford Mustang Mach-E - the full stack from perception to lateral & longitudinal C++ control over a drive-by-wire CAN bus on an NVIDIA Jetson AGX Orin, with nonlinear acados MPC control for lateral path-tracking and the ACC, smooth enough for stop-and-go.
The perception I own is RADAR + Vision fusion: my RADAR point-cloud tracker paired with classified vision detections into one reliable set of tracks. I also built a self-improving camera pose-estimation module that turns raw drives into clean odometry ground truth.
That builds on a research foundation in the CAVAS Lab at University at Buffalo (Autoware bring-up on a Lincoln MKZ, ROS2 control, and multi-sensor fusion), and before autonomy, two years in industrial automation where real-time, embedded, safety-critical thinking became second nature. I'm now looking for planning roles where decisions meet the road.
02
What I build
One driving loop: a real on-car clip, plus live views of sensing the scene, planning the behavior, and putting it on the road.
Real on-car UI, straight from the vehicle
The exact on-car display: fused RADAR + Vision, the locked lead with live range and speed, the planned path, and real-time speed with the forward-camera inset.
Perception · Radar
4D-radar tracking
My 4D-radar multi-object tracker (Kalman with Doppler updates, 18 Hz) uses vision-guided lane assignment, lifting correct far-lead selection in curves from 54% to 87%; a zero-phantom stopped-vehicle detector confirms in-lane stopped vehicles at ~100 m.
Path-Relative
Planning
Planning is path-relative: distance s runs along the model's path, lateral offset d across it - along s I shape the speed (gap-keeping, stop-and-go, curve-speed); the lateral MPC holds d.
Lat MPC + Long IDM
Control
A real-time acados nonlinear MPC handles lateral control (the model's path to curvature, roll-compensated feedforward plus gated PI trim); longitudinal control is IDM car-following with time-headway gaps, a stop-and-go / crawl / launch state machine, cut-in handling and lead arbitration.
Training side: how the label is made (Hybrid A*)
On-car: the model predicts the path in construction zone directly from camera
03
Experience
- Integrated Autoware.Universe on a Lincoln MKZ connected autonomous vehicle (Ubuntu 22.04, ROS 2 Humble) and stood up the C++ drive-by-wire control stack executing Autoware's planned trajectories.
- Authored the vehicle and sensor-kit launch and calibration configurations, bringing up the full multi-sensor perception and localization pipeline.
- Implemented LiDAR-IMU extrinsic calibration (Ceres) and LIO-SAM factor-graph SLAM with RTK-GNSS factors (~2 cm fixes, geodetic→ENU), building a georeferenced HD point-cloud map of campus for map-based localization.
- Led the full-stack autonomy project: containerized simulation and control modules with Docker and mentored 2 students building a digital twin.
- Prototyped a V2V module broadcasting ego state for cooperative speed coordination and remote takeover between connected vehicles.
- Delivered 10 embedded automation systems integrating PLCs with multi-axis CNC motion for profile cutting, including closed-loop automatic torch-height control.
- Configured industrial PCs and Eckelmann PLCs for real-time servo control over CANopen and EtherCAT; wrote CODESYS PLC logic and G-code on an E°EXC 882 controller.
- Built custom HMIs and led firmware updates with servo-drive vendors to synchronize motion across all axes.
04
Technical stack
Languages & Core
C++PythonC
Real-time systemsVehicle dynamicsState estimation
Planning & Control
acados MPCPure pursuitAdaptive Cruise Control
Hybrid A*RRT / RRT*Dubins pathsDrive-by-wire
Perception & Fusion
RADAR + Vision fusionObject trackingRADARSensor fusion
Camera pose estimationLiDARGNSS / RTK
Platforms & Tooling
ROS2 (Humble)Autoware.UniverseNVIDIA Jetson AGX Orin
CAN / CANopen / EtherCATDockerUbuntu 22.04CodeSYS / PLC
05
Projects
Production · IDrive
IDrive Autonomy - Mapless L2+ ADAS
End-to-end self-driving on a Ford Mustang Mach-E - perception, planning and C++ actuation on a Jetson AGX Orin, no prior map required.
Perception → Planning → Control
Perception
RADAR + Vision Fusion Tracking
My RADAR point-cloud tracker, fused with classified vision detections, locking onto every vehicle around the car even in heavy sensor noise.
Downstream of vision-detection inference
Control
acados MPC Path Tracking
Lateral and longitudinal control that holds the model's reference path at highway speed.
C++ · on the vehicle
Research · CAVAS Lab
Autoware on a Lincoln MKZ
Autoware.Universe bring-up on a ROS2 connected autonomous vehicle - multi-sensor fusion and a C++ control stack over the drive-by-wire bus.
University at Buffalo
06
Get in touch
Let's put planning on the road.
If you're building a planner or driving-behavior stack and want someone who takes it from decision to actuation, I'd love to talk.
dkgosai@ymail.com