Clinical-grade metrology. Cloud processing. Any camera.
Built for imaging cores, metrology teams, and instrument builders who need quantitative phase and 3D surface data they can defend in an audit — without buying a quarter-million-dollar interferometer.
Pilot labs report reconstruction quality at clinical level and higher SNR than the internal code they were running — on homemade and open-source instruments they built because a factory-calibrated scope costs $250K.
Every precision manufacturer faces the same bottleneck: measurement hardware that costs more than the production line it monitors.
Dedicated interferometers with 12–18 month lead times and single-vendor lock-in.
Proprietary software, annual calibration contracts, and black-box algorithms.
Per-line cost that scales linearly with every new manufacturing cell.
If you already have holograms — or a camera and a laser — you already have everything you need to start.
You have a DHM rig, or a camera pointed at an interference pattern, and you are writing reconstruction code instead of running experiments. Send the hologram; get back amplitude, unwrapped phase, and a focus curve.
Why it matters: stop maintaining a private reconstruction pipeline, and get a number you can cite — NCC 0.88 against published Zenodo UTIRnet ground truth.
You need surface data on the line, not in a lab six weeks out. Replace the processing back-end of a $250K instrument with an API call, and turn an existing machine-vision camera into a calibrated measurement station.
Why it matters: the same input and config returns a bit-for-bit identical result, every time — the audit-trail story no ML-based system can tell.
You ship optical hardware and carry a reconstruction stack you would rather not maintain. Embed ours — or run our independent third bus as a quality gate on the pipeline you already have.
Why it matters: on-premises deployment, no GPU dependency, and a per-result provenance package your customers' auditors will accept.
Three steps from raw hologram to quantitative surface data.
Point any industrial camera at a laser interference pattern.
Send the image via HTTP POST with optical parameters.
Get back micrometer-accuracy 3D surface data in seconds.
An optical-domain bus, an LPU dual-path bus, and a statistical bus each reconstruct the field. Pixels where they disagree are rejected by MAD outlier vote at 2.5σ.
Disagreement between the three buses is the noise model. Wiener filtering is driven by measured per-pixel variance, not an assumed noise floor.
28-mode Zernike fit (IRLS, Noll ordering) strips residual tilt, defocus, coma, and spherical from your optics — then TV-ROF regularisation, gated by a Tamura texture guard so real edges survive.
Every result ships a quality_report.json: phase residue count, consensus agreement score, per-bus noise estimate, and a provenance hash.
Shoot one hologram of a known object — an M1.6 hex nut for off-axis, a prepared salt-crystal slide for inline — and our digital-twin optimiser solves for your true pixel pitch, wavelength, sensor tilt, and aberration map. Save the profile once; every later reconstruction from that camera is corrected automatically.
Live at api.apertureapi.com. Free tier, no sales call, no SDK to install — anything that can POST multipart form data works.
curl -sS -X POST https://api.apertureapi.com/users/signup \
-H "Content-Type: application/json" \
-d '{"email":"you@example.com","password":"…","plan":"free"}'
# → { "access_token": "eyJhbGci…", "token_type": "bearer",
# "user": { "tier": "free", "api_key": "ak_…" } }Free tier includes 100 reconstructions/month. Paid plans add a Stripe payment_method_id at signup.
curl -sS -X POST https://api.apertureapi.com/reconstruct_pdhm \
-H "Authorization: Bearer $TOKEN" \
-F "file=@hologram.tiff" \
-F "wavelength_um=0.635" \
-F "pixel_pitch_um=1.85" \
-F "holography_mode=off-axis" \
-F "quality_mode=max_snr" \
-F "deterministic=true" \
-o reconstruction.npz
Defaults are the production path.
quality_mode=max_snr runs the full consensus pipeline; deterministic=true pins seeds and BLAS threads for reproducible output.
Switch holography_mode to inline for Gabor geometry (GS-20, HIO+ER, or HIO+ER+shrinkwrap twin-image suppression), or auto to detect it from the Fourier carrier ring.
| NPZ key | Shape | Contents |
|---|---|---|
amplitude |
(H, W) | Reconstructed amplitude field |
phase |
(H, W) | Unwrapped phase, radians |
focus_distance_mm |
scalar | Autofocus result |
focus_curve |
(N,) | Per-z focus metric trace |
metadata |
JSON | Full parameter set + calibration profile ID, for reproducibility |
Accepts TIFF, PNG, JPG, NPY, NPZ, and HDF5 holograms. The /reconstruct endpoint returns a ZIP visualization package instead — 9-panel analysis PNG, PLY and OBJ 3D models, rotation video, and raw NPZ.
| Endpoint | Auth | Purpose |
|---|---|---|
POST /reconstruct_pdhm |
Bearer | Production reconstruction. Off-axis, inline, or auto-detected. |
POST /reconstruct |
Bearer | Reconstruction with a full visualization package (ZIP). |
POST /autofocus |
Bearer | Find the focal plane without a full reconstruction. |
GET /calibration/targets |
Bearer | Catalog of reference objects, with buy links and CAD ground truth. |
POST /calibration/calibrate |
Bearer | Fit a calibration profile for your camera. Returns a profile_id. |
POST /public/reconstruct |
None | Try it without an account. HMAC-nonce gated, free-tier quota. |
POST /users/signup |
None | Create an account and mint a bearer token. |
GET /health · /docs |
None | Service health and interactive OpenAPI reference. |
Physics-based reconstruction engineered for production environments.
Physics-based reconstruction with auditable, repeatable results. No machine learning, no training data, no black box.
Three independent reconstruction buses cross-check every measurement, with MAD outlier rejection at 2.5σ. 99.9% combined error detection; 0.912 average consensus agreement.
One hologram of a $5 reference object solves for your camera's true pixel pitch, wavelength, tilt, and aberration map. Commodity hardware, calibrated optics-grade results.
35–50ms per frame for core reconstruction, 85–130ms for the full pipeline, at 1024×1024 on a CPU. No GPU required. Temporal streaming kernel hits 60fps at 512×512.
Photonic and analog optical processors are fast because they compute in the light itself — and unreliable for exactly the same reason. Sensor noise, drift, and aberration mean a result often has to be run several times before anyone trusts it. There is no deterministic verification layer for optical computation. We built one, and pointed it at metrology first.
Three independent propagation reconstructions vote on every pixel. Disagreement is not discarded — it is the measured noise model. That is a verification primitive, not a filter.
A patent-pending reversible transform maps the optical domain into the acoustic one, letting 60 years of mature audio DSP act as a fully independent third opinion on a light-field result.
No model weights, no training set, no inference. Same input and config returns a bit-for-bit identical result — the property an optical compute node needs before anyone will certify it.
The reconstruction kernel runs today on CPU, with GPU and FPGA paths in progress. The same kernel that verifies a hologram verifies an optical compute result — which is why we treat metrology as the beachhead, not the ceiling.
Deployed across sectors where measurement accuracy defines product quality.
Solder paste inspection, BGA measurement, microvia depth profiling. Inline and offline configurations.
Wafer defect detection, etch depth metrology, MEMS profiling. Process control at nanometer resolution.
Surface roughness measurement, micro-lens array characterization, scratch/dig quantification.
Label-free cell imaging, drug response monitoring, blood diagnostics. FDA-suitable deterministic processing.
If you run a digital holographic microscope, an interferometer, or a rig you built yourself — send us your raw sensor data and we will reconstruct it at no cost. No purchase order, no license, no instrument to replace. We want hard samples and honest feedback; you get reconstructions you can put in a paper.
Raw holograms, your wavelength and pixel pitch, and what you are trying to measure.
Amplitude, unwrapped phase, 3D surface files, and a per-result provenance report.
Tell us where it fails. Difficult samples improve the calibration for everyone.
Email mzeligs@alumni.stanford.edu — goes straight to the founder, not a queue.
Replace six-figure capital equipment with a software subscription.
Zygo, KLA, Keyence, and Lyncee Tec all sell the same answer to this problem: a better camera, factory-calibrated, at a six-figure price. We moved the moat to software — which is why a 3D-printed rig plus our API produces results a lab will publish.
| Vendor | System Cost | Annual Maint. | ApertureAPI | You Save |
|---|---|---|---|---|
| Zygo (Ametek) | $150K–300K | $15–30K/yr | $10K setup + API | 93% |
| KLA-Tencor | $300K–2M | $50–200K/yr | Enterprise API | 95%+ |
| Keyence | $40K–120K | $5–12K/yr | $10K setup + API | 70–85% |
| Lyncee Tec | $100K–500K | $10–15K/yr | API access | 98% |
Based on replacing equivalent legacy system at average inspection volume.
Predictable pricing that scales with your inspection volume.
Free tier: 100 reconstructions/month, no card required. Start in one curl command →
Pay-as-you-go: $1/scan, no commitment.
All tiers include ApertureAPI docs, multi-region availability, and SOC 2 compliance.
MULTI-GRANULARITY VERIFICATION FOR OPTICAL PROCESSORS WITH SPATIAL LOCALIZATION.
Mathematical Method and System for Transcoding Between Electromagnetic and Acoustic Spectra.
DUAL-PATH OPTICAL SIGNAL PROCESSING SYSTEM WITH ADAPTIVE FUSION.
Start a free lab pilot, read the API docs, or open the investor data room.
Founded by Michael Zeligs — Stanford CCRMA, signal processing; physics awards from the University of Colorado; imaging and photography awards since age fourteen; fifteen years building cameras and imaging companies. The insight behind ApertureAPI came from the audio side of the house: phase and amplitude recovery is the same problem in light that it is in sound, and the optics industry kept answering it with better hardware instead of better math.