Synthgen: Product, team and how it was built
Synthgen (Synthetic labelled images of rare defects for training vision models, Netherlands): Finalist in the AI Pitch Competition 2026, September 2026. Evidence grade E: no revenue figure was checked. Team: Started late 2024 as a student project; co-founder and CTO Martin Tomov; a second co-founder (go-to-market) joined in October 2026; headcount not published, careers page says "small team". Launched 4 December 2024; still running.Added 6 November 2026
- Outcome
- Finalist in the AI Pitch Competition 2026, September 2026
- Evidence
- E no revenue figure was checked
- Launch
- 4 December 2024
- Status
- Still running
- Team
- Started late 2024 as a student project; co-founder and CTO Martin Tomov; a second co-founder (go-to-market) joined in October 2026; headcount not published, careers page says "small team"
- Outside funding
- Funded: the company said OWOW Venture Studio backed it with €200,000 in capital and services as a pre-seed (October 2026), and that it received a €124,000 grant from the Netherlands Enterprise Agency (November 2025)
- Main channel
- Not stated
What is not verified
No revenue figure was found. No source read gives a customer count, a named paying customer or a headcount; the careers page says its products are "used by computer vision teams in production" without naming any. The €200,000 pre-seed and the €124,000 grant rest on the company's own blog posts. The headline is the finalist announcement of 18 September 2026 on the company's blog, which IO+, an organiser, confirms. A reader could look for the KVK filing of the company (90725859), the result of the 12 November 2026 final, or a customer case study. This story is grade E: the business is real and its product, price and customers are described from public pages, but no revenue figure was checked. Aiolum shows it as an idea worth studying, not as proof that it makes money.
Synthgen, a company in Eindhoven, the Netherlands, sells synthetic data: images made by a generative model that look like real camera frames, used to train another model. Its customers run computer vision models, programs that detect or classify objects in images, on inspection lines in factories and farms. The weak point of such a model is the rare class, a defect type with very few training examples. Synthgen generates more images of that class and adds automatic labels, the boxes and pixel masks that say where the defect is, which people normally draw by hand.
The company's about page says it was founded in late 2024 as a student project, and its blog says it launched officially on 4 December 2024 at a pitch evening on the High Tech Campus Eindhoven. IO+ names Martin Tomov as CTO and co-founder. On 18 September 2026 the company announced that it was one of eight finalists of the AI Pitch Competition 2026, whose final is on 12 November 2026. In October 2026 the company said OWOW Venture Studio backed it with €200,000 in capital and services.
No revenue figure, customer count or named customer was found, so this is a grade E story: the headline is a competition result. The company publishes its own benchmark, which it calls SCARCE, with gains of 9 to 35 percentage points on the hardest class of each of nine public datasets.
How much revenue does Synthgen make?
Each row shows the evidence grade and the source numbers from the list at the end of this page. The grades are explained in How Aiolum verifies.
| Fact | Value | Grade | How we know | Sources |
|---|---|---|---|---|
| AI Pitch Competition 2026 | one of eight finalists; final on 12 November 2026 | C | Company blog, 18 September 2026: "Eight startups came through a competitive field to reach the final of the AI Pitch Competition 2026, and Synthgen is one of them." IO+, one of the organisers, profiled the finalists on 2 November 2026. | [6, 1] |
| Founded | late 2024, as a student project | C | About page: "Synthgen was founded in late 2024. We started it as a student project, kept shipping after graduation, and turned a capstone into a company." | [3] |
| Official launch | 4 December 2024 | C | Company blog, 4 December 2024: launched at the 100th Demos, Pitches & Drinks at the High Tech Campus Eindhoven. | [8] |
| Founders and team | CTO and co-founder Martin Tomov; a go-to-market co-founder joined in October 2026; headcount not published | C | IO+, 2 November 2026: "CTO and co-founder Martin Tomov". Company blog, 19 October 2026: a co-founder for go-to-market joined from OWOW. The careers page says "Hard problems, small team". No source gives a headcount or names the other founders. | [1, 5, 12] |
| Pre-seed | €200,000 in capital and services | C | Company blog, 19 October 2026: OWOW Venture Studio is "backing us with €200,000 in capital and services and a multi-year partnership". The split between cash and services is not given. | [5] |
| Grant | €124,000 from the Netherlands Enterprise Agency | C | Company blog, 9 November 2025, on its Bloomberg TV Bulgaria interview: the grant "supports our work with Luxeat Robotics", a Dutch maker of laser weeding robots. The site footer also says "Funded by the Dutch MIT R&D scheme (RVO), 2025–2027". | [7, 2] |
| Benchmark result | +9 to +35 percentage points on the hardest class | C | Research page: across 9 datasets and 8 industries, "7 to 15 seeds, on a frozen test set", paired against real-data-only baselines. Example: pill imprint defects 46% to 81%. The company ran and published the benchmark itself. | [10, 2] |
| Prices | not published; quoted per Assessment and per image | C | Pricing page, read 6 November 2026. | [4] |
| Target markets | food and produce lines, precision agriculture, electronics and precision manufacturing | C | Home page and about page: "We sell first into food and produce lines, precision agriculture, and electronics and precision manufacturing." | [2, 3] |
| Company register | KVK 90725859, Eindhoven | C | Site footer. The register entry itself was not read. | [2] |
| Time from launch to the finalist announcement | 21 months | calc | Calculated: December 2024 to September 2026. | [8, 6] |
How to read this table: a grade C value comes from the company's own pages or from the IO+ article. A "calc" value was computed on this page. No value is grade A or B, and no revenue figure exists.
| Plan | Per month, billed yearly (US$) | Per month, billed monthly (US$) | Limit |
|---|---|---|---|
| Synthetic Data Assessment | one month, fixed scope, quoted; compute included | ||
| Per image, pay as you go | price per generated and labelled image, quoted; no subscription | ||
| Monthly packs | prepaid, lower price per image for larger packs, quoted | ||
| Enterprise | custom terms |
The pricing page, read on 6 November 2026, says: "Why are there no prices on this page? Because a published number would be wrong for most people who read it." The Assessment is quoted after scoping meetings, and images are quoted by volume and resolution. Labels are included in the image price.
When did Synthgen launch, and how did it grow?
From a student capstone to a pitch-competition final with a published benchmark and a venture studio behind it, in about two years. Each entry carries its grade.
- Late 2024CSynthgen starts as a student project. Its first goal, by the company's account, was to generate wildlife images for AI training; it then moved to agriculture. [3, 7]
- 4 December 2024COfficial launch at the 100th Demos, Pitches & Drinks at the High Tech Campus Eindhoven. [8]
- 1 February 2025CJoins the SPARC incubator program and moves into an office at TQ, Strijp-S, Eindhoven, with a link to Fontys ICT. [9]
- 9 November 2025CThe company reports a €124,000 grant from the Netherlands Enterprise Agency for work with Luxeat Robotics, and an interview on Bloomberg TV Bulgaria. [7]
- 18 June 2026CPublishes SCARCE, its first public technical report: a placebo-controlled benchmark across nine datasets. [10, 11]
- 18 September 2026CAnnounced as one of eight finalists of the AI Pitch Competition 2026. [6]
- 19 October 2026CThe company says OWOW Venture Studio backs it with €200,000 in capital and services; a go-to-market co-founder joins. [5]
Why did Synthgen work?
Four causes, in the order they mattered. Each one is a reusable lesson.
Technical depth
The product is a generative model tuned on a customer's own frames, plus a benchmark method to prove that it helps. The about page calls the company "research-native" and names diffusion models, zero-shot detection and synthetic augmentation as its field. That is work a small team of machine-learning graduates can do without a factory of its own.
Labels by construction
When a program places the defect in the image, it already knows where the defect is, so it writes the bounding box and the segmentation mask at the same time. Manual labelling, often the largest cost of a vision project, disappears for the generated images.
Proof before volume
The paid Assessment freezes a held-out set of the customer's real images before anything is generated, then compares the model trained with and without synthetic data on that set. A buyer who has heard many AI promises sees a number on their own data before buying images at volume.
How it was built
These details come from the company's own pages. They are grade C. [2, 11, 3]
- The Synthgen Engine learns the rare classes from a customer's images; a new class is "generated and labelled the same day", according to the home page.
- Output includes bounding boxes, segmentation masks and classification labels, delivered in the customer's training format.
- Datasets are imported from and exported to Hugging Face, Weights & Biases and Amazon S3.
- Trained models can run in a data centre or on a compact edge device.
- The site publishes an llms.txt file, an MCP server for agents and an OpenAPI description of its site search and contact endpoints.
- The site's content-security policy allows connections to RunPod, a GPU cloud, and to Amazon Web Services; no source names the generative model used.
Tools named in the sources
Each one opens its page in the Tools tab, with every story that uses it.
Who tried this before Synthgen, and why did they fail?
Three companies in our database of failed startups sold advanced technology to factories or ran research as a business. The values below are from that database and are not independently verified.
Rethink Robotics
The promiseCollaborative factory robots for about US$25,000 that workers could train by demonstration.
Why it diedThe robots were about four times slower than industrial arms, so customers did not get their money back.
Daqri
The promiseAugmented-reality helmets that gave factory workers instructions and sensor data.
Why it diedHeavy, expensive helmets, clumsy software and factory customers who were not ready to buy.
Argo AI
The promiseOne self-driving system that car makers could buy instead of building their own.
Why it diedIts backers decided that the work was open-ended research without predictable business milestones.
The failed startups come from our database of 1,749 dead startups, retold here in our own words. Its entries are AI-assisted summaries of public sources, so these cards carry no evidence grade.
What changed
Rethink Robotics and Daqri sold new hardware to factories and failed on cost and on slow adoption. Argo AI shows the other risk of a research-led company: years of work without a date when customers pay.
Synthgen sells no hardware; its home page says "No camera, no appliance" and "Keep the cameras and the vision stack you already run". Diffusion models, which the about page names as its field, and rented GPUs let a small team generate realistic defect images. The one-month Assessment is meant to give a customer a measured result early. Whether the Assessments turn into paying volume is not yet in any source.
Can one person copy Synthgen now?
Hard for one person; it needs machine-learning research skill.
The method is public and the tools are open, but the value is in a generator that works on many kinds of images and a benchmark that buyers trust. A founder with computer-vision research experience and contacts in one industry could start the same way, with one class of defect in one kind of line.
Where it is still open: Other rare-event vision problems: medical-device and pharmaceutical inspection, X-ray and other non-destructive testing, security screening, and weeds or crop diseases for farm robots.
- 1Pick one industry where you can get real images and where a rare defect is costly.
- 2Fine-tune an open image-generation model on a dozen real examples, and write labels when the defect is placed.
- 3Publish a benchmark on public datasets that includes the cases where synthetic data did not help.
- 4Sell a short paid test on the customer's own held-out images before selling images at volume.
What could go wrong for Synthgen?
The risks are the mirror image of the causes above.
What is not verified
No revenue figure was found. No source read gives a customer count, a named paying customer or a headcount; the careers page says its products are "used by computer vision teams in production" without naming any. The €200,000 pre-seed and the €124,000 grant rest on the company's own blog posts. The headline is the finalist announcement of 18 September 2026 on the company's blog, which IO+, an organiser, confirms. A reader could look for the KVK filing of the company (90725859), the result of the 12 November 2026 final, or a customer case study.
A service step before the software
Most customers start with an Assessment in which Synthgen's engineers audit the data, generate samples and write a costed plan, ending with "a readout call with the engineer who built it". The image generation is done by the company's own engine and sold per image; the site has a sign-in, but no self-service checkout takes a payment. The product is software output with a consulting-like first step.
Self-published benchmark
The +9 to +35 point gains come from the company's own SCARCE report on public datasets. The company says it includes classes where synthetic data did not help. No independent replication was found.
Competition and model providers
No other story on this site sells synthetic training data; the closest is TEMS.AI, a Dutch two-founder company that sells AI software to factories. No source names the image-generation model; the site's security policy allows connections to RunPod and Amazon Web Services, so GPU rental prices reach the product's costs directly.
Sources
The grade next to each source is the best grade that the source can support. "Read in full" means that the text of the page was downloaded and the numbers were checked in it.
- C IO+, 2 November 2026: Nine AI startups, one final: from smart promises to real valueRead the section on Synthgen. Source for the co-founder's name and the final on 12 November 2026.
- C Synthgen home page, read 6 November 2026Read in full. Source for the product, target markets, integrations, the RVO funding line and the register number.
- C Synthgen About page, read 6 November 2026Read in full. Source for the founding date and the student-project origin.
- C Synthgen Pricing page, read 6 November 2026Read in full. Source for the Assessment and the per-image plans.
- C Synthgen blog, 19 October 2026: Synthgen raises €200,000 pre-seed from OWOW Venture StudioRead in full. Source for the pre-seed and the new co-founder.
- C Synthgen blog, 18 September 2026: Synthgen is a finalist in the AI Pitch Competition 2026Read in full. Source for the headline.
- C Synthgen blog, 9 November 2025: We Got Interviewed by Bloomberg TV BulgariaRead in full. Source for the grant, the Luxeat Robotics work and the early pivot. The TV interview itself was not watched.
- C Synthgen blog, 4 December 2024: Synthgen Debuts at High Tech Campus EindhovenRead in full. Source for the launch date.
- C Synthgen blog, 1 February 2025: Synthgen Joins SPARC Incubator ProgramRead in full. Source for the incubator and the office.
- C Synthgen Research page, read 6 November 2026Read the summary of results. The PDF report was not downloaded.
- C Synthgen llms.txt, read 6 November 2026Read in full. Source for the list of blog posts, the agent endpoints and the product description.
- C Synthgen Careers page, read 6 November 2026Read in full. Source for the "small team" statement.
- C Database of failed startups, local exportLocal export. AI-written summaries, not independently verified. Source for the failed predecessors.
Run this company? Send corrections or a verified figure to aiolum@livyden.com