I'm a mechanical engineer who works where hardware meets software. At Tesla I build test systems and autonomous robots. After hours, I built and shipped HiddenCloak alone.
The glowing root is my real route, Rourkela to San Jose, on true compass bearings from home. Signals travel it in the shape of a crash pulse. Each cap is a chapter; hover one.
At Tesla I designed and integrated a system that drives a real Tesla through the car's internal firmware: two GPS antennas and a separate base station, a Raspberry Pi running Kalman fusion, Stanley steering and PID speed control. It replaced a $450k system for about $2.5k.
Scroll, and the drawing comes apart into the four disciplines it took.
I learned engineering by breaking cars properly. At Maruti Suzuki I reworked seat and body structures until they passed crash regulations. At Tesla I still draft parts in 3DX with GD&T.
Sensors, compute and the wiring between them. My autonomous test system rides on a real Tesla: two GPS antennas on the roof, a separate base station sending corrections over radio, and a Raspberry Pi that talks to the car's internal firmware. I build test hardware myself when buying it makes no sense.
Closed loops. At Tesla I designed and integrated an autonomous test system that drives a real Tesla through the car's internal firmware. Two GPS antennas and a separate base station feed a Kalman filter on a Raspberry Pi; the Stanley law steers it along the path and PID holds its speed. It replaced a $450k system.
Tools people actually use. Python suites that check hundreds of crash channels, a model that flags bad sensors after a test, and the Raspberry Pi controller that drives the car, and HiddenCloak end to end: React Native, Express, Postgres, and two AI providers judging each other.
# the steering law from the Lab simulation (illustrative, not Tesla code)
def stanley(path, pose, v, k=1.2):
i = path.nearest(pose.front_axle)
heading_err = wrap(path.yaw[i] - pose.yaw)
cross_track = path.signed_dist(pose.front_axle, i)
delta = heading_err + atan2(k * cross_track, v + 0.4)
return clip(delta, -MAX_STEER, MAX_STEER)
I started in a crash lab, where the job is to break a car properly and learn exactly how it failed. Four years later, I moved to Texas for a master's degree, then to Tesla, where the test rigs started driving themselves. Scroll down and the forest floor travels the route back with you, city by city, to the roots in Rourkela.
HiddenCloak · iOS, live on the App Store
Designed, built and shipped a consumer AI app on my own: the mobile app, backend, AI pipeline, moderation, App Store review and website.
Back to the bloom
Tesla · Fremont, California
I own sensing system performance from prototype to deployment: integration test plans, vehicle-level bring-up, and characterization across environments. I also debug where hardware and software meet.
Tesla · Fremont, California
I brought machine learning and automation to post-crash data, so engineers spent their hours on findings instead of spreadsheets.
Texas A&M University · College Station, Texas
Controls, vehicle systems and simulation. Graduate projects modelled in Modelica and Dymola: an EV dual-battery pack, a hybrid powertrain controller and an engine-start study (see Field work).
Maruti Suzuki India Ltd, R&D · Rohtak, Haryana
Four years of crash compliance and research for India's largest carmaker, across AIS, ECE and NCAP programs. I took new and existing vehicles from design prototype release through testing within four months.
National Institute of Technology · Rourkela, Odisha
The proprietary hardware stays at work. These are working sketches of the ideas behind it, so you can poke at them.
At Tesla I designed and integrated an autonomous system that drives a real Tesla over a ditch jump, holding a straight (or nearly straight) line into the ramp. It fuses GPS through a Kalman filter, steers with the Stanley law, holds speed with PID, and commands the car through its firmware over CAN and Etherloop. It replaced a $450k system for about $2.5k.
Each run here starts off the line, and Stanley pulls it in before the lip. Drag the nodes up or down to bend the approach, and change the gain to see how the controller trades responsiveness for overshoot.
δ = ψe + arctan(k·e / v)
A simulation of the published control law, not Tesla code.
Crash tests need impact speed you can trust. Two laser gates 1000 mm apart, a Raspberry Pi timing the breaks, and a Python UI give v = d / Δt for about a tenth of the commercial price.
Simulated runs at typical frontal-test speeds.
A crash test produces hundreds of sensor channels, and one bad one can hide in the noise. A random forest trained on previous tests labels every channel good or bad right after the test, shown in a GUI, saving 75% of the man-hours.
Synthetic traces for illustration.
IIHS's updated moderate overlap test puts a 5th percentile female dummy in the rear seat. A belt that rides up the chest makes the sternum gauge read low, so the chest index scales the measured deflection by where the belt sat at the moment of maximum deflection. My Python scripting turns those channels into this kind of animation, so the rating can be read and checked, not just reported.
Synthetic traces. Method and boundaries (35 / 40 / 45) from the IIHS Moderate Overlap 2.0 Rating Guidelines (Version II, June 2023). Belt positions above 110 mm are rated on belt load and position instead.
What I reach for, from crash labs to app stores. Hover a tool and it lights up in its own colours.
English, Hindi, Odia, German
Best Employee Award · Best Graduate Project · Best design award, trolley transfer method
Autonomy, renewable energy, robotics, music, working out, reading
Happy to talk robotics, mechatronics, test systems, vehicle safety, or HiddenCloak.