Sr. Principal Systems Engineer · AEye
Physical AI · Sensing · Real-Time Perception
I build the sensing and early perception layer for Physical AI: sensors that decide, in real time, where to look next based on what they just saw, and hand a machine the world model it acts on.
Founding-team engineer at AEye. I joined before the first raise in 2016, carried seven generations of LiDAR through to a $1.5B IPO, and I am still writing the production code a decade later.
[email protected] · linkedin.com/in/joelbenscoter · Gilbert, Arizona, open to relocation
Pre-funding to IPO, and still shipping
The most recent ten years of a twenty-three year career, and the chapter most of this page is about. Each round closed on a capability that had to work first. The full history is further down.
Joined before the first raise. Product definition and first-customer engagement.
Delivered the AE10 prototype, demo vehicle and visualizer, and fulfilled the first customer contract in full.
Led the iDAR, 360°, 100 fps and 1 km demonstrations that closed the round.
2,000+ pages of RFI/RFQ response, converting into NRE contracts and design wins with major automotive OEMs and Tier 1s.
Launched 4Sight M mid-pandemic with ~200 remote demos in one month. $260M closed.
Final technical authority for sensor architecture. Still hands on the code every day.
What it moved
Six results, before any of the detail. Each is the sensor’s own measured performance or a business outcome that followed from it, not a projection. The rest of this page is how they were built.
Delivered the NVIDIA 300 m result, published with the highest points-per-second on NVIDIA's sensor list, and the announcement that followed.
230K to over 8M points/sec, through shot scheduling, the three-shot processor, upscaling, wire compression and SIMD.
8 hours to 15 minutes to under 5 seconds. A simulation that was scheduled overnight is now interactive.
34.7 seconds to 9.9. Six root causes, including 21 million redundant register reads and a six second sleep that fired on every cold boot.
Technical owner of the eye-safety evidence under 21 CFR 1040.10 and IEC 60825-1 Ed. 3.
End to end validation of ADC pixel summing against a suspended golf ball 7.8 dB below the noise floor, the result that de-risked the defense and infrastructure programs.
Six 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 is a different discipline with a different toolchain. I work all six, which is the part most engineers don't. The full record, measured from 1,655 commits →
Where the laser fires, 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 a year across automotive, Intelligent Transportation Systems, rail, defense and calibration.
Full-frame ADC captures pulled off the sensor, with the FPGA's matched-filter and CFAR datapath reproduced in simulation. How the CFAR end-of-buffer bug, the MSMS waterfall bug and DC switching noise were found.
The embedded post-processing pipeline. Dynamic Range Coincidence, near-range retro/bloom suppression, low-SNR, spatial and transparency filters, sustaining more than 8 million points per second.
The 3D viewer engineers and customers actually look at, plus the offline harnesses that grade filter changes against recorded ground truth instead of opinion.
Named technical lead on the programs that carry the company: a publicly announced NVIDIA engagement, plus automotive OEMs, defense primes, rail operators, off-road autonomy and national labs. The other four stages get pointed at whatever the customer's hard problem is that quarter.
Technical owner of the eye-safety analysis behind the FDA/CDRH Class 1 certification, and of the 11-stage signing pipeline that makes every shipped scan pattern auditable.
The algorithms that decide where to look and what is real
It starts before the photon: where the laser fires is itself an optimization. After that, a raw return is not a point. Between the photon and the point cloud sits the work that separates a real object from a reflection, a noise spike, rain, or the bloom off a road sign. This is the layer perception teams inherit, and most of it is mine.
Five constraints that fight each other. A shot can only fire when the mirror is pointing at that azimuth, on one of two sweep branches, far enough after the last one for the laser to recharge, inside the usable swing, and still hitting the field of view, resolution and frame rate the customer asked for. None of it is linear: mirror position is sinusoidal in time, laser recovery exponential. Solved exactly rather than by heuristic: a bitmask dynamic program that maximizes the minimum inter-shot gap, a Jonker-Volgenant assignment pairing up-sweep against down-sweep, and a Frantz-Nodvik recurrence predicting what each pulse will really deliver. Greedy, then Round Robin at 2×, then Turbo at 3×.
Coincidence across neighboring returns, so a real surface survives and an isolated artifact does not. Four generations: v1 a static threshold, v2 peak width, v3 multi-dimensional slope support, v4 refactored for high speed embedded compute. Concept through production, and the layer downstream perception teams inherit.
Every echo carries four bits for the direction it failed to find a coincident neighbor, west, east, south, north, plus interior corner and ground plane bits. That is the object outline, labeled per point, before any perception stack sees the cloud. Coincidence is tested against the local range slope rather than a flat range match, so a car angled away or mid turn stays one surface instead of shattering into edges, and the ground is separated from what stands on it.
A retroreflector at close range washes out everything around it. This finds the retro core, flags the bloom, and keeps the neighbors.
Pulls real returns out of the noise floor at long range, where the target is a few photons above nothing.
Removes isolated noise without eating the small distant object that is the whole reason for a long range sensor. Found its low signal to noise retention path had never once executed in a shipping configuration.
Multi row handling for glass and mesh, so the sensor reports the surface and what is behind it rather than picking one.
Was deleting the edges of trees near retroreflectors. Proved across 3.7 million shots that peak width alone cannot tell real structure from artifact, then added a neighbor veto that spared 28% of those points with no threshold loosened.
Four returns from one shot, so rain, dust and foliage do not hide the thing behind them, and weak returns a two echo limit would discard still reach the point cloud. That last part is what recovers low signal to noise targets.
Physics based, overlapping beams beyond the laser’s full width half maximum with confirmation shots. 9× the throughput and 1.5× the range out of the same laser, which is how the sensor reaches targets the shot rate alone cannot.
The detection layer under all of it. Re-scaled an encoding that was silently crushing 4% of returns to a single value, measured across 20.7 million returns, and traced a three year old filter bug to it.
Telling two close things apart instead of merging them into one.
Summing across ADC pixels to lift a target out from beneath the noise floor, where a single pixel has nothing to report. Carried from the theory through the FPGA datapath model to a measured result: a suspended golf ball detected at 208 m, 7.8 dB below the noise floor, which de-risked the defense and infrastructure programs.
The engineering record
I pulled 33 months out of 14 repositories and charted it: what I work on, which codebases I own, how the tempo moved, and what it produced. It is the part of a resume you normally have to take on faith.
Fifty-one customer programs, six markets
Every scan pattern is cut for one customer's optics, range, resolution and timing budget. The library I maintain covers fifty-one distinct programs, the subset of the customer base that needed a pattern built specifically for it. Most are under NDA, so they are grouped here by market. The partners AEye has announced publicly are named.
Passenger car, ADAS and windshield integration programs across US, European, Japanese, Korean and Chinese manufacturers. Publicly announced: GM, Continental, LG, LITEON.
Primes, autonomy integrators, national laboratories, a government research agency, and space and eVTOL programs. Publicly announced: SynTech.
High speed rail, urban transit, signalling and autonomous freight across Chinese, French and European operators, plus airport ground safety, perimeter detection and fiber sensing. Publicly announced: Black Sesame Technologies.
Highway and intersection monitoring, gantry tolling, traffic intelligence and roadside deployment. Publicly announced: Blue-Band, Flasheye, Vueron, Michigan DOT, Mitsubishi Electric (MEAA), Alive 3D.
Driverless fleets and the compute platforms they run on. Publicly announced: NVIDIA and the GM sponsored all weather autonomy research program at the University of Toronto.
Long haul autonomy, electric freight and heavy equipment operating outside mapped road networks.
Counted from the customer pattern library.
The engineer in the room, before and after the sale
For most of a decade I have been the technical subject matter expert customers actually talk to. Not a sales engineer handed a deck, but the person who wrote the scan pattern, the filter and the firmware, answering for it directly. That shortens the loop: what a customer needs in the morning can be in a pattern by the afternoon.
A real requirement, in the customer's words
Six points on a 10 × 50 cm, 10% reflective object at 160 km/h closing speed on the autobahn.
Pre-sales
Post-sales
Scan pattern to partner perception
01
I write the pattern to the requirement. Where the shots go, how often each region is revisited, how much energy each one gets, and what that costs in frame rate and eye safety budget.
02
Field capture on the customer's own scene, with their mounting geometry, their range and their weather. The pattern gets revised against what came back, not against a simulation.
03
Their classifier or tracker runs on those points. Vueron highway monitoring is a public example of the far end of this loop.
The sensor is not the deliverable. What a partner's perception stack can do with the points is, and that only works if the pattern was designed for their problem in the first place. Owning all three steps is what makes the turnaround short.
Twenty-three years, six employers
AEye is the current chapter and the longest one, but the habits came from somewhere: seven years shipping semiconductor inspection hardware, a startup I co-founded and wound down, and a decade of infrastructure work before either.
Seven generations of 1550 nm LiDAR, pre-funding through a $1.5B IPO. Six promotions in nine years, 57 patents, and the whole signal path. Everything above.
Datacenter delivery for the SaaS storage, archiving, e-discovery and compliance portfolio. Hardware refresh programs across four datacenters, 112 customers and 1,000+ systems. Co-led a $60M contract.
Consumer photo management startup, built alongside the KLA-Tencor role. Grew to 102 customers and 350,000 photos processed with a team of nine. Funded the beta with a crowdfunding campaign that raised 114% of goal and trended first in its opening day. Chose to wind it down.
Semiconductor wafer inspection. Architected server and network systems processing over 15 billion pixels per second, roughly half a petabyte an hour, for a $15M product that booked $1B in its first year. Shipped 37 embedded products across the full life cycle and led the company's first multi-division product.
Five years of IT and network infrastructure before the move into product engineering.
The evidence
Three pages, with the full 57-patent portfolio as an addendum.
15 artificial intelligence · 11 machine learning and software perception · 11 adaptive hardware · 11 feedback controls · 3 hardware · 2 functional safety · 2 system controls · 2 manufacturing and compliance. Core inventions: iDAR, dynamic shot scheduling, low-latency motion planning, upscaling, boresighted lidar/camera, bi-static architecture, compressive sensing.
Currently open to new work
Principal, staff and distinguished engineer roles, or chief architect. LiDAR and optical sensing, real-time embedded systems, signal processing, or anything where the hard part sits between the physics and the software. Happy to talk through any of the work above in detail.
Private portfolio Engineering figures and tooling, shared on request. Password required.