Electrical & Computer Engineering · USC · Class of 2028
AKSHAY
VADLAMANI
I build things that have to keep working when something goes wrong. Rocket flight software,
chips that repair their own memory, and perception that runs on the robot instead of the cloud.
Looking for Summer 2027
NASA JPL · Summer 2026
Los Angeles
Where I've worked
Jun 2026 — Aug 2026 · Pasadena, CA
NASA Jet Propulsion Laboratory
Research Intern · Analysis & Test Group
Spacecraft are starting to run neural networks. Nobody had a good answer for what happens
when a cosmic ray flips a bit inside one. I spent the summer building the equipment to find out.
- Built a fault injection platform in Python, corrupting weights, activations, and instruction
streams across nine campaigns on an NXP i.MX93 (Arm Ethos-U65 NPU) and a Microchip
PolarFire SoC (RISC-V U54).
- No int8 runtime exists for RISC-V, so I wrote a TensorFlow Lite interpreter from scratch in
NumPy — flatbuffer parsing, operator traversal, quantized convolution and softmax — and
validated it bit-exact against a TensorFlow reference across 100 images.
- Found that the model compiler, not the silicon, drives how the two platforms fail. That puts
the compiled model inside the hardware qualification envelope, where nobody was looking for it.
- The worst failures are silent. Monitoring caught 97 of 97 hardware refusals and 0 of 10 wrong
answers, separated by 0.008 ms of latency against a 0.023 ms standard deviation across 439
runs. A confidently wrong inference looks exactly like a correct one.
PythonNumPyTensorFlow LiteRISC-VEmbedded LinuxremoteprocLibero
Jan 2026 — Present
USC Liquid Propulsion Lab
GNC Software Engineer
- Vertical state estimator for a liquid rocket flight computer: a 3-state Kalman filter in C++
on Zephyr RTOS fusing 400 Hz VN-300 inertial data with GNSS and LiDAR altimeter updates,
Joseph-form covariance, innovation gating.
- Redundant-sensor fault detection across a 1000/400/50 Hz stack. The estimator used to freeze
on a sensor dropout; now it degrades per channel so navigation survives.
- Throttle control firmware on ARM Cortex-M7 with 1 ms hard-real-time loops. Trig goes into
lookup tables because transcendental latency is data-dependent and doesn't belong in a
control path.
C++Zephyr RTOSCortex-M7KalmanProtobuf
Jan 2026 — Present
Orbes
Robotics & perception · Techstars-funded
- Camera-based localization for autonomous indoor drones on a Jetson Orin Nano, working with
Intel RealSense stereo depth and RGB-D as the front end for visual SLAM.
- Ultrasonic short-range rig covering the near-field blind zone the cameras can't see.
- The product won $10K at TroyLabs Demo Day & Pitch Night.
Jetson Orin NanoRealSenseVisual SLAMArduino
2023 — 2024
VR Safety Research
Research Intern · NSF-funded, $850K
- Built VR environments in Vizard with Python scene scripting for infrastructure
emergency-response training.
- Spatial algorithms for 3D asset placement and collision detection, plus a pipeline turning
real campus infrastructure survey data into usable 3D models.
PythonVizard3D spatial
Things I've built
Sep 2025 — Dec 2025
Kairo
Form analysis for judged sports
Diving, gymnastics, and figure skating are scored on technique that coaches assess by eye.
Kairo measures it instead.
- Vision stack on MediaPipe BlazePose pulling 33 3D skeletal landmarks per frame, with
three-point joint angles computed in 3D and visibility gating that returns nothing rather
than a wrong number when a limb is occluded.
- Electron and three.js desktop app rendering motion capture against recorded video on a
frame-synced timeline. Currently mid-migration to TypeScript.
- Six collegiate programs paid for it in a single semester, including Caltech diving and
athletics at Stanford, Oregon State, USC, Redlands, and Berkeley.
TypeScriptElectronthree.jsMediaPipeComputer vision
→ github.com/detco20256/kairo-frontend
Jun 2026
Radical
Radiation-hardened INT8 CNN accelerator
An inference chip that assumes its own memory is lying to it.
- Four-layer conv/FC pipeline in SystemVerilog behind AXI-lite and AXI-stream, with SECDED
Hamming ECC on the weight datapath, triple-modular-redundant MAC units, and a background
scrubber that walks weight memory correcting upsets before they're read.
- Chip-level fault injection testbench proving single-bit upsets get corrected with the output
class held, and double-bit upsets get flagged uncorrectable instead of silently accepted.
SystemVerilogSECDED ECCTMRAXIVerilator
Sep 2025 — Present
Aeropet
Gesture-controlled drone · USC Makers
- MediaPipe hand landmarker models deployed through Qualcomm AI Hub, running 21-point hand
tracking on the NPU onboard rather than streaming to a server.
- Gestures have to hold across consecutive frames before they become flight commands, which is
what makes it usable instead of twitchy.
- MQTT publish/subscribe pipeline moving camera frames and commands in both directions at
720p/30fps, over a Rubik Pi configured with UART, SSH, and SCP.
Qualcomm NPUTFLiteMQTTOpenCVUART
Dec 2025
APEX
Embedded ML · reps in reserve
- Arduino Nano BLE Sense streaming 3-axis IMU at 100 Hz through a first-order high-pass filter
at 0.1 Hz, because integrating raw acceleration means integrating gravity too.
- Trapezoidal integration for velocity, threshold-based rep segmentation, eight features per
rep into a Random Forest that predicts how many reps you have left before failure.
- Won Best Product in EE 105, Ming Hsieh Department of Electrical Engineering.
ArduinoLSM9DS1scikit-learnSignal processing
Oct 2025
Mobius
Natural language to 3D math
- Next.js and Plotly.js frontend rendering parametric 3D surfaces with camera path
interpolation.
- FastAPI backend using OpenAI function calling to parse plain-language math, classify equation
types, and return visualization parameters; Math.js handles symbolic processing and
coordinate generation.
Next.jsPlotly.jsFastAPIMath.jsRailway
In progress
INT8 CUDA Convolution
Kernel optimization on a T4
- Six kernel versions from a naive NCHW baseline through shared-memory tiling, NHWC layout,
and
__dp4a vectorization, each isolating exactly one change so the speedups are
attributable.
- The interesting part is the roofline. Arithmetic intensity is 527 ops/byte and the whole
working set fits in the T4's 4 MB L2, so nothing here is DRAM-bound, including the naive
version. Tiling is fighting L1/L2 traffic, not memory bandwidth.
- Measurements pending on hardware. Numbers go here when they're real, not before.
CUDANsight ComputeRooflineQuantization