Computer Vision Engineer

Why are we hiring this role?
We are building a system that turns a bid package into quantities someone can price. This role owns
“what is it?” — identifying the features, whether they are drawn on a plan or visible from the air. The
other engineering role we are hiring owns “where is it?” — geometry and measurement. The two meet
at the quantity, and getting from one to the other is shared work.
Two input paths, both yours. Drawings arrive in wildly mixed condition — some carry usable geometry,
others are scanned, flattened or degraded and need real computer vision. Aerial and satellite imagery is
the second path: the same corridor seen from above, where the job is detecting surfaces and features
that the drawing may not show or may show wrongly. The two have to agree, and where they disagree
that is itself a finding.
Automated takeoff is well covered for buildings. Heavy civil highway work is a harder problem and a
much thinner field — that is the one you would own.
What you’ll own
- Routing what to look at. Most of a bid package is prose — specifications, proposals, addenda —
and only a minority of pages carry geometry. The same is true of imagery. Deciding what
deserves compute before spending it is what keeps this affordable.
- Extraction from drawings — vector and raster. Linework out of PDF vector content streams,
plus the computer vision path for sheets where that data is missing or unreliable: skew
correction, thresholding, line and contour extraction on inputs that are rotated, faded, or
rescanned.
- Identifying the material, and the items around it. The question a contractor is actually asking is
what material, and how much of it. Asphalt or concrete. Which mix, which layer, what thickness,
what base under it. From drawings that means reading pavement sections and material callouts
alongside the geometry — plus drainage structures, inlets, pipe runs, edges, lane lines, signs and
guardrail, with text and annotation tied to what it describes. From aerial imagery it means
telling asphalt from concrete on the real ground, and finding the extent of each — pavement
area, lane breakdown, driveways, approaches. Both paths resolve to the correct pay item,
including where the drawing, the imagery and the specification disagree.
- The pipeline around the models, and the quantities out of it. Labeling tooling, dataset
versioning, an eval harness, and regression monitoring, built from scratch. Detections have to
become quantities someone will price, carrying confidence and provenance per item — which
means working shoulder to shoulder with the geometry side rather than throwing output over a
wall.
What we expect day to day
- You prioritize completeness over cleverness. A wrong number usually gets caught, because the
total looks wrong. A missing one does not, and it reaches the submitted bid. We would rather
flag fifty items as not determinable than silently drop three.
- You return “not determinable” instead of a low-confidence guess. And you record what was
inferred rather than read. CAD exports print attributes but not object identity or network
membership, so relationships have to be reconstructed — and marked as reconstructed.
- You own labeling personally. Ground truth gets built here, alongside the people who know the
domain. This is not a clean-data environment and there is no labeling team to hand it to.
- You work in the open and take feedback straight. We are direct. You ship progress to the team
channel, present at reviews, and get honest feedback — including from the founder. We argue
about the work, decide fast, and move.
- You use AI aggressively. Everyone here uses AI daily for research, writing, analysis, and building.
Same bar for this role.
What We’re Looking For
We only hire superstars — expect to be asked why you’re one.
Required
- 5+ years shipping production CV/ML systems, with end-to-end ownership from labeling through
deployment on at least one of them
- Strong Python and PyTorch, with production shipping — not research prototypes
- PDF internals — content streams, text layers, path construction, and why a drawn line is often
many segments — or the appetite to learn them fast.
- Computer vision on real imagery: segmentation and detection on aerial, satellite or comparable
raster data — resolution limits, tiling, and what is genuinely recoverable at a given ground
sample distance
- Classical CV for degraded scans: OpenCV, thresholding, line and contour extraction, template
and feature matching
- Comfort with sparse, messy domain data and human-in-the-loop systems
- Strong written and verbal English; clear async communication
Nice to have
- Deep specialty in either half — aerial and satellite CV, or document AI and OCR on engineering
drawings, floor plans and CAD vectorization. Strength in one and working knowledge of the
other is the realistic shape.
- Backgrounds we love: aerial and geospatial imagery, drone mapping, construction or document
CV, or autonomous-vehicle perception. Extraction runs as a standalone Python service; the
platform it feeds is TypeScript, and experience there is not required.
- Anything 3D — photogrammetry, point clouds, terrain and surface models, cut and fill. Surfaces
are how the geotechnical and earthwork side gets reached, and it is the natural next thing after
material and area. Not required, but we would rather hire someone who can already see that
far.