Ph.D. · Algorithm Team Lead, Neolithics.ai · Israel
Ahmad Droby
I build computer-vision systems that grade fresh produce from RGB and 224-band hyperspectral imagery, and I study how vision models learn from less supervision.
Reflectance spectrum · 400–1000 nm · 224 bands
590 nmband 72 of 224ripe 0.06unripe 0.12bruised 0.16
ripeunripebruised
Schematic reflectance of fruit tissue across the range a VNIR hyperspectral camera sees: the green peak of chlorophyll, the red edge near 700 nm, the water band at 970 nm. Illustrative shapes, not measured data. Move across the plot to read a band; the amber tick marks 590 nm, the wavelength of the A above.
Now
Role
Leading the algorithms team at Neolithics.ai: multimodal RGB + hyperspectral models that grade blueberries, avocados, grapes and garlic, from dataset lifecycle to TensorRT on edge hardware.
Research
Weakly-supervised and unsupervised learning, historical-document analysis, and dynamic 3D Gaussian splatting. Latest: Incoherent Deformation, Not Capacity (2026), on why dynamic scenes overfit and how to stop it.
Before
CTO and co-founder of Mirage Dynamics (AI-driven in-video advertising, 2022–2024). Ph.D. from Ben-Gurion University, 2023, on weakly-supervised learning for visual computing.
Selected work
Eight papers I would hand a new collaborator first: text-line and page segmentation with little or no supervision, Hebrew and Arabic paleography, and the dynamic-splatting diagnostic.
- 2026
- 2022
- 2022
- 2022
- 2020
- 2020
- 2018
- 2017
All 21 publications, 2017–2026
- 2026
- 2022
- 2022
- 2022
- 2022
- 2022
- 2022
- 2022
- 2022
- 2021
- 2021
- 2021
- 2021
- 2021
- 2021
- 2020
- 2020
- 2020
- 2018
- 2018
- 2017
Also on Google Scholar, DBLP, OpenAlex and ORCID. Per-paper citation counts are OpenAlex's, which indexes fewer citing works than Scholar.
Experience
2025 – now
Algorithm Team Lead · Neolithics.ai
Own the computer-vision and hyperspectral roadmap for automated fresh-produce quality control, and the team that ships it.
- Designed and lead the model build flow: data collection, dataset lifecycle, training and deployment as one reproducible pipeline.
- Multi-label and ordinal heads (CNN, ViT and hybrids) with severity-aware losses, for defects that come in degrees rather than classes.
- Deployment standards: TorchScript / ONNX / TensorRT, config-logged training runs, versioned datasets on DuckDB, GCP and zarr.
2024
Senior Computer Vision Researcher · Neolithics.ai
A multimodal fusion network blending RGB with 224-band hyperspectral data for real-time quality prediction; end-to-end pipelines from ingestion to automated validation; Jetson-ready hybrid CNN / transformer inference.
2022 – 2024
CTO & Co-founder · Mirage Dynamics
Led the technology for AI-driven advertising inside existing video: detection, segmentation and tracking to replace on-screen ads, dynamic insertion for live HLS streams, contextual analysis with captioning and zero-shot classification. Co-inventor, with Jihad El-Sana, on the company's patent application for video advertising signage replacement (WO 2021/090328).
2018 – 2023
Doctoral researcher & teaching · Ben-Gurion University
Visual Media Lab, advised by Prof. Jihad El-Sana: weakly-supervised learning for page layout, text-line segmentation and script classification on historical manuscripts. Taught and assisted computer-science courses; Teaching Excellence Award, 2022.
Projects
Research
Incoherent Deformation, Not Capacity
A diagnostic study of overfitting in dynamic Gaussian splatting: the failure is incoherent per-Gaussian deformation, not model size, and it can be mitigated. Paper, figures and code on the project site.
idnc.drobya.com ↗Product
Wird
A calm Quran companion for tracking a khatma: daily portions, progress and gentle reminders. Web first, iOS in progress.
wird.drobya.com ↗Lab
Ayn Research
An experimental lab where a team of autonomous agents generates, critiques and scores ideas in visual intelligence, and I keep the ones that survive. Private, for now.
Home
Solar & home-automation lab
A 25 kW inverter, a three-phase meter and a Home Assistant hub feeding power-quality analysis: the same data-engineering discipline, pointed at my own grid connection.
Education & awards
2018 – 2023
Ph.D. in Computer Science
Dissertation: Using Weakly Supervised Learning to Solve Visual Computing Problems. CHE doctoral scholarship, 2019–2022; Hi-Tech & Bio-Tech entry scholarship, 2018.
2016 – 2018
M.Sc. in Computer Science
Thesis: Surface Detection and Deformation Detection in a Video Stream. Academic Excellence Award, 2018; CHE master's scholarship.
2013 – 2016
B.Sc. in Computer Science
Awards
- Teaching Excellence Award, Ben-Gurion University, 2022.
- 1st place, page-segmentation task, RASM2018 competition at ICFHR 2018 (Ben-Gurion University team).
Contact
Open to collaboration on computer-vision research, hyperspectral imaging and production ML for industrial inspection.
LinkedIn Google Scholar GitHub ORCID 0000-0001-8458-1022 DBLP OpenAlex Curriculum vitae PDF