Engineering footprint · Jan 2024 to Aug 2026

From photons to pixels, one engineer.

Joel Benscoter works the entire LiDAR signal path: designing where the laser fires, modeling the energy it carries, simulating what the FPGA sees, writing the real-time C++ that turns returns into points, building the tools that make those points legible, and sitting across the table from the customer who buys them. Very few people own more than one link in that chain. This is what owning all five looks like in the commit record.

Measured from 14 repositories at AEye. Source files only. Captured sensor data, spreadsheets and binaries are excluded.

1,655
Commits authored
351K
Lines of code added
1,208
Source files touched
5
Codebases, 3 languages

The work

Five links in one signal path

A LiDAR return starts as a scheduled laser shot and ends as a colored point on someone's screen. Each stage below is a different discipline with a different toolchain, and each one shows up in the same person's commit history.

01

Scan patterns & radiometry

MATLAB & C++ · where the laser fires, and how hard

Designs the shot schedule itself: which direction, in what order, at what energy, inside eye-safety and thermal limits. Sole author of the company's scan-pattern output: 42 new patterns in 2025 across automotive, Intelligent Transportation Systems, rail, defense and calibration. Rewrote the eye-safety simulation to model the full transmit chain, and owns eye-safety signing. Took the radiometry model from an 8-hour run to 15 minutes, and then to under 5 seconds, fast enough to sweep configurations interactively instead of scheduling them overnight. Automated pattern-to-sensor install from 15 minutes to roughly 20 seconds.

Shot schedulingEye safetyLink budget Beam & parallaxHex codegen 207 commits tagged
02

FPGA & raw capture

Hawkeye · what actually comes back

Built the host-side tooling to pull full-frame raw ADC captures off the sensor and reproduce the FPGA's matched-filter and CFAR datapath in simulation, which is how the CFAR end-of-buffer false-positive bug, the MSMS waterfall bug and DC switching noise were found. Proved the ADC pixel-summing theory end to end against a suspended golf ball at 208 m, 7.8 dB below the noise floor.

CFAR modelingMatched filtersGain channels Noise floor48 commits tagged
03

Real-time sensor service

C++ · returns become points, at 8M points per second

The embedded post-processing pipeline. Invented and productized Dynamic Range Coincidence; built near-range retro-bloom suppression, the low-SNR pipeline, the unified spatial filter, transparency and squelch filters, and the quad-echo path. Cut point-cloud streaming from 40 bytes per shot to 11 bytes per point, added SIMD through the hot loops, and rebuilt buffering so the sensor sustains more than 8 million points per second.

DRCNR-RBMLow SNR Spatial filterSIMDWire protocol Top contributor of 12
04

Client, visualization & test harnesses

Qt / C++ · making a million points per frame legible

The 3D viewer engineers and customers actually look at: filter visualization for every stage of the pipeline, a point-debug inspector that dumps a single return's neighborhood, PCAP replay, colormaps and range controls, an in-app profiler. Alongside it, the offline harnesses: regression testing with capture provenance, and a retro-bloom grader that scores filter changes against recorded ground truth instead of opinion.

3D renderingPoint debugPCAP replay Regression harness234 commits tagged
05

Customer programs

The reason the other four exist

Named technical lead for the accounts that carry the company. Fifteen-plus pattern revisions for NVIDIA to hit 300 m at 10% reflectivity with the highest points-per-second on NVIDIA's own published list, the one engagement that is public, and the announcement behind a material liquidity event. Beyond it: automotive OEMs and Tier 1s, defense primes, rail and transit operators, off-road autonomy, infrastructure monitoring and a national laboratory. The pattern work, the filters and the visualizer all get pointed at whatever the customer's hard problem is that quarter.

NVIDIAAutomotive OEMsDefense primes Rail & transitOff-road autonomyNational labs 113 commits tagged

Ownership

Not a contributor. The contributor.

Share of all source lines added to each codebase across the full window, measured against every other author on the repository.

Share of lines added, by codebase

Jan 2024 to Aug 2026 · all branches · source files only

The MATLAB modeling and scan-pattern toolchain is essentially a one-person codebase. On the shared C++ sensor service, a twelve-engineer repository, he still leads on volume.

Domains and languages

What the commits are about, and what they are written in

Every commit message classified by subject matter. A single commit can carry more than one tag. A scan-pattern change made for a customer counts in both. Below that, the same body of work seen from the other side: which language it landed in.

Commits by technical domain

1,655 commits, 14 domains

Fourteen distinct technical domains in one person's history: optics, signal processing, real-time systems, graphics, and the customer work that drives all of them.

Languages, by lines added

A modeling language and a systems language, in near-equal measure

The split is the point: prototype the algorithm in MATLAB where the math is cheap to change, then ship it in C++ where it has to run in real time on the sensor. In 2026 that loop closed: the MATLAB toolchain was ported to a native C++/Qt application, 67,000 lines in seven weeks, with a parity gate that checks the C++ against MATLAB goldens.

Tempo

Thirty-three months, none of them empty

Every month from January 2024 through September 2026 has commits in it, and 38% of them land before 2026. The 2026 acceleration is one identifiable event: the C++ port of the modeling toolchain and the DRC v3 architecture arriving together.

Commits, cumulative

Every month adds to the line. The slope is the rate.

Thirty-three consecutive months, no gap anywhere in them. Hover any point for that month’s running total and what it added.

Outcomes

What it moved

The code is the evidence; these are the results it produced. Each figure below is the sensor's own measured performance, not a projection.

Sensor throughput

Points per second sustained, Dec 2023 → 2026

A 35× increase, driven by the Turbo 3 scheduling algorithm, the three-shot processor, upscaling, wire-format compression and SIMD, five separate pieces of work across three codebases.
$80M+
Liquidity unlocked

The NVIDIA 300 m result and its public announcement extended the company's runway.

45×
Faster pattern turnaround

Scan-pattern install cut from 15 minutes to ~20 seconds, enabling 100+ variations and closing a four-year intermittent flicker bug.

5,760×
Faster radiometry

Full-system modeling cut from 8 hours to 15 minutes to under 5 seconds. A simulation that was scheduled overnight is now interactive.

−72%
Streaming bandwidth

Point-cloud wire format from 40 bytes/shot to 11 bytes/point, fitting ~9.5M points per second inside 1 GbE.

57
Patents

Issued and pending, spanning AI, adaptive hardware, feedback controls and perception.

208 m
Golf ball, at the noise floor

End-to-end validation of ADC pixel summing, the result that de-risked the defense and infrastructure programs.