Overview
OnDeck AI is an active Vancouver-based startup founded in April 2022 by Alexander Dungate and Sepand Dyanatkar. Originally incorporated as OnDeck Fisheries AI, the company built computer vision software to automate fish biomass counting and species identification from shipboard video — replacing costly human observers in commercial and regulatory fisheries monitoring. It deployed in the US, Canada, and Costa Rica, attracted support from National Geographic, Schmidt Marine Technology Partners, and the Government of Canada, and won a $1.5M grant from Canada's Ocean Supercluster in October 2023.
This is not a post-mortem. OnDeck is a live company that used a niche, high-friction vertical as a paid R&D environment, validated a generalizable technical insight, and deliberately pivoted to a larger market before the original vertical could constrain it.
In January 2025, having reached approximately $150K ARR in fisheries, the founders pivoted to a general-purpose enterprise vision infrastructure platform — one that lets organizations analyze any footage without training a model. By October 2025, the company had grown from $0 to six-figure ARR post-pivot, secured a $160K pilot with the Singaporean Military, and joined Y Combinator's Summer 2025 batch. The fisheries chapter was not a failure; it was a deliberate proving ground that stress-tested the technology in conditions most AI vision startups never encounter.
Founding Story
Alexander Dungate and Sepand Dyanatkar have been close friends since grade six — a founding dynamic that is either a strength or a liability depending on how conflict gets handled, though in this case the partnership has held through three years of pivots and co-founder departures.[1]
Dungate studied Computer Science and Biology at the University of British Columbia, where coursework on global fisheries management surfaced what he described as "egregious, existential problems" in the industry.[2] He became a National Geographic Explorer and developed a specific thesis: fisheries monitoring was a rate-limiting step for sustainable fisheries management globally, and "nothing's changed in 20 years."[3] In 2021, he pitched the idea to Dyanatkar and a third friend, Matthew Leighton, who would become CFO.
Dyanatkar brought the technical depth the idea required. He holds a Master's in machine learning (specializing in multi-agent reinforcement learning) from Cambridge University, and had worked at the European Space Agency and Amazon before co-founding OnDeck.[4] His background in swarm robotics and distributed ML systems would later prove directly relevant to the company's VLM decomposition architecture.
The three incorporated OnDeck Fisheries AI in Vancouver in April 2022.[5] The name was deliberate — domain-specific rather than platform-oriented, signaling a vertical focus at inception. Within months, the company was winning startup competitions: first place at the University of Washington's Dempsey Startup Competition ($25K USD) in May 2022, and first place in Canada's Ocean Idea Challenge ($10K CAD) in August 2022.[6]
By fall 2022, the founding team had contracted. Matthew Leighton stepped back to return to his PhD at Simon Fraser University.[7] The terms of his departure — including any retained equity — are not publicly disclosed. Dungate and Dyanatkar continued as a two-person founding team, hiring their first software engineer and assembling an advisory board that included Dr. Daniel Pauly and Dr. Rashid Sumaila, two of the most cited fisheries scientists in the world.[8]
The initial vision was explicitly domain-specific: build the electronic monitoring infrastructure that global fisheries regulators needed but couldn't afford to deploy at scale. The pivot to general-purpose enterprise vision infrastructure in January 2025 was not a repudiation of that vision — it was a recognition that the underlying technology had outgrown the vertical that produced it.
Timeline
- 2021 — Alexander Dungate pitches fisheries AI concept to Dyanatkar and Leighton after UBC coursework identifies systemic problems in global fisheries management[2]
- April 2022 — Company incorporated as OnDeck Fisheries AI in Vancouver, BC[5]
- May 27, 2022 — Wins first place ($25K USD Herbert B. Jones Grand Prize) and $5K Glympse IoT Prize at University of Washington's 25th annual Dempsey Startup Competition[6]
- June 14, 2022 — GeekWire covers OnDeck Fisheries AI, describing the ML/CV product for automated fish biomass tracking[2]
- August 4, 2022 — Wins second place ($5K USD) at WashU Olin Big IdeaBounce; wins first place ($10K CAD) in Canada's Ocean Idea Challenge; named to New Ventures BC Top 25[9]
- Fall 2022 — Matthew Leighton (CFO, third co-founder) steps back to return to his PhD at SFU[7]
- October 4, 2023 — Awarded $1.5M in direct funding from Canada's Ocean Supercluster for the "AI for Scalable Fisheries Monitoring Project," with total project value of $3.5M including matching funds[10]
- February 1, 2024 — First software engineer publishes Medium post documenting early team structure and advisor network[7]
- December 4, 2024 — em4.fish profiles OnDeck Fisheries AI, noting deployments in US, Costa Rica, and Canada[11]
- January 2025 — Company pivots away from fisheries monitoring to general-purpose enterprise vision infrastructure, having reached approximately $150K ARR in the fisheries vertical[12]
- April 14, 2025 — VanTech Journal profiles Alexander Dungate, documenting the pivot rationale and CAD $170K+ in prior grants and awards[3]
- May 2025 — Post-pivot ARR growth begins from $0; company begins accumulating enterprise customers across defense, robotics, security, and oil & gas verticals[13]
- August 20, 2025 — YC S25 launch post published: "OnDeck AI — Analyze any video without training a model"[14]
- August 25, 2025 — American Bazaar reports on YC S25 batch, noting OnDeck has analyzed thousands of hours of footage across autonomous surface vehicles, robotics, security, and oil & gas[15]
- October 27, 2025 — Creative Destruction Lab profile documents $0 to six-figure ARR since May 2025, $160K Singaporean Military pilot, and YC + National Geographic + Eric Schmidt foundation backing[13]
What They Built
Phase 1: Fisheries Monitoring Software (2022–2024)
The original product addressed a specific operational problem in commercial and regulatory fisheries: verifying what fish are caught, in what quantities, and of what species — without requiring a human observer on every vessel.
Traditional fisheries monitoring relies on government-employed observers who ride commercial fishing boats, manually counting and recording catch data. This is expensive, logistically difficult, and impossible to scale globally. Electronic monitoring — cameras mounted on vessels — had existed for years, but the footage generated far more data than human reviewers could process cost-effectively.
OnDeck's Phase 1 product was computer vision software that automated this review. Cameras mounted on fishing vessels captured footage of fish being brought aboard; OnDeck's models identified species and estimated biomass from that footage. The system was designed to replace or substantially reduce the human review burden, making electronic monitoring economically viable at scale.[16]
The core technical challenge was data scarcity. As Dungate explained: "The challenge with computer vision models is that they need a lot of training data; hundreds of thousands of images."[17] Fisheries monitoring involves hundreds of species, many rarely photographed in commercial catch contexts, across wildly varying lighting conditions (shipboard, underwater, night operations). Building a traditional supervised learning model for this domain would require labeled datasets that simply did not exist.
The team's response to this constraint — rather than attempting to collect the missing training data — was to investigate whether Vision Language Models (VLMs) could generalize to niche species identification tasks without domain-specific training. They published a NeurIPS workshop paper demonstrating that VLM-based methods outperformed traditional computer vision benchmarks on species identification, even for rare or niche species.[12] This finding was the technical pivot point that would eventually reshape the entire company.
The fisheries product was deployed in the United States, Costa Rica, and Canada through partnerships with governments, monitoring companies, NGOs, commercial fishers, and Indigenous communities.[11] By January 2025, it had reached approximately $150K ARR.[18]
Phase 2: Enterprise Vision Infrastructure (2025–present)
The current product is an infrastructure layer that makes Vision Language Models accessible and scalable for enterprise customers — delivered via a developer-friendly API.[12]
The flagship offering is called Perception-0, described on the company's website as "the world's most powerful vision model, guaranteed to maintain SOTA over time."[19] The technical approach involves decomposing VLM components and recombining them into a general-purpose visual reasoning system — rather than deploying a single monolithic model.[20]
The key differentiator versus traditional computer vision platforms is the absence of a training data requirement. Customers do not need to label images, collect domain-specific datasets, or fine-tune models. This matters most in three scenarios: (1) regulated or sensitive environments where sharing training data creates legal or security risk; (2) novel or rare event detection where labeled data doesn't exist; and (3) qualitative behavioral analysis — detecting sequences of events or contextual behaviors rather than just classifying objects.[21]
The product is SOC 2 certified and can be deployed in cloud, on-premises, or air-gapped configurations — a deliberate architectural choice that opens defense and critical infrastructure markets that cloud-only competitors cannot serve.[19]
By summer 2025, the system had analyzed thousands of hours of footage across autonomous surface vehicles, robotics research platforms, security camera networks, and offshore oil and gas monitoring installations.[15]
Market Position
Target Customers
OnDeck's post-pivot customer base clusters around a specific profile: organizations that generate large volumes of video footage in operationally sensitive or data-scarce environments, where traditional CV fails and where the cost of a missed detection is high.[12]
Documented customer verticals include national defense organizations (including the Singaporean Military), robotics research companies, security camera operators, port monitoring operations, and offshore oil and gas installations.[12] The Singaporean Military use case — finding targets and behaviors in footage from a fleet of autonomous surface vessels — is the clearest illustration of the product's positioning: high-stakes, novel event detection, in an environment where training data is classified or nonexistent.[22]
The company is also deployed with universities and public safety organizations,[19] though these appear to be secondary to the defense and critical infrastructure anchor customers.
Market Size
The enterprise video analytics market is large and growing. Estimates vary widely by scope, but the broader computer vision market is projected to exceed $25 billion by 2030 across industrial, security, and defense applications. The specific segment OnDeck targets — training-data-free video analysis for regulated and sensitive environments — is a subset of this, but one where pricing power is substantially higher than in commodity CV deployments. The $160K Singaporean Military pilot alone suggests contract sizes that are an order of magnitude above typical developer API consumption.[13]
The defense and critical infrastructure segment is also structurally attractive: procurement cycles are long, but contracts are sticky, renewal rates are high, and switching costs are significant once a system is integrated into operational workflows.
Competition
OnDeck competes across two distinct dimensions, and its position on each is worth mapping carefully.
On the infrastructure/API dimension, the company competes with general-purpose computer vision APIs: AWS Rekognition, Google Video Intelligence, and Microsoft Azure Video Analyzer. These incumbents have massive distribution advantages — they are already embedded in enterprise cloud contracts — but they are fundamentally training-data-dependent systems. They excel at detecting objects from categories they were trained on; they struggle with novel behaviors, rare events, and qualitative reasoning. OnDeck's no-training-data architecture is a genuine differentiator here, not a marketing claim.
On the enterprise AI platform dimension, the company competes with Scale AI (data labeling and model evaluation infrastructure), Palantir AIP (AI integration for defense and government), and a range of specialized defense CV vendors. Palantir is the most relevant competitive threat in the defense vertical: it has deep government relationships, existing security clearances, and a proven enterprise sales motion. OnDeck's advantage is technical agility and the absence of the organizational overhead that makes Palantir deployments expensive and slow.
The structural competitive question is whether the major VLM providers — OpenAI, Google DeepMind, Anthropic — will build native video analysis APIs that commoditize OnDeck's core offering. This is a real risk. However, the company's moat is not the VLM itself; it is the decomposition and recombination architecture that makes VLMs operationally deployable in air-gapped, latency-sensitive, and security-constrained environments. General-purpose VLM APIs are unlikely to offer SOC 2 certified, air-gapped deployment with enterprise SLAs in the near term. The fisheries origin story also provides a differentiated narrative: the founders have real-world deployment experience in adversarial conditions — ships, low-light environments, novel species — that stress-tested the technology in ways that lab-built competitors have not matched.
Business Model
OnDeck has not publicly disclosed its pricing structure. Based on available evidence, the company appears to be pursuing an enterprise contract model rather than high-volume developer API consumption — a reasonable inference given the customer profile (defense agencies, oil and gas operators, port authorities) and the $160K Singaporean Military pilot contract size.[13]
The company's Phase 1 fisheries business reached approximately $150K ARR before the January 2025 pivot.[18] The Phase 2 business grew from $0 to six-figure ARR between May and October 2025 — a roughly five-month ramp that implies either a small number of large contracts or a rapid accumulation of mid-sized pilots.[13] The Singaporean Military pilot at $160K is consistent with the latter interpretation.
Inferred burn rate (labeled as inference): With 5 employees listed on YC's directory[23] and open roles at $100K–$250K CAD,[24] a rough annual burn estimate of $800K–$1.2M CAD is plausible, assuming fully-loaded costs of $160K–$240K CAD per employee. At six-figure ARR (conservatively $200K–$400K USD), the company is not yet cash-flow positive, but the trajectory from $0 to six figures in five months suggests the gap is narrowing.
The company has never disclosed revenue figures directly. The $150K ARR (fisheries) and "six-figure ARR" (post-pivot) figures both originate from founder-sourced claims on the YC and CDL profiles — not audited financials.
Traction
OnDeck's post-pivot traction is the most concrete signal available about whether the enterprise vision infrastructure thesis is working.
Between May and October 2025 — five months post-pivot — the company grew from $0 to six-figure ARR, plus a $160K pilot with the Singaporean Military.[13] The "six-figure ARR" claim is founder-sourced and unaudited, but the specificity of the Singaporean Military pilot contract value ($160K) lends credibility to the broader traction claim.
By summer 2025, the system had analyzed thousands of hours of footage across multiple verticals: autonomous surface vehicles, robotics research, security systems, and offshore oil and gas monitoring.[15] The breadth of verticals in under six months suggests the no-training-data value proposition is resonating across use cases rather than being confined to a single niche.
YC acceptance for the S25 batch (August 2025) represents institutional validation of the pivot thesis.[14] YC's selection process is competitive, and the S25 batch acceptance — coming after the pivot rather than during the fisheries phase — suggests the admissions committee evaluated the enterprise vision infrastructure thesis on its own merits.
The company is simultaneously enrolled in Creative Destruction Lab,[13] a Canadian deep-tech accelerator with a strong track record in AI companies. Dual enrollment in YC and CDL is unusual and suggests the founders are actively managing their network and mentorship resources.
Key unknowns: customer churn rates, contract lengths, renewal rates, and whether the Singaporean Military pilot has converted to a recurring contract. These are the metrics that will determine whether the six-figure ARR figure represents durable revenue or a collection of one-time pilots.
Post-Mortem
This section is structured differently from a standard post-mortem because OnDeck is not defunct. Instead, it analyzes the strategic logic of the Phase 1 fisheries business — why it was built, what it accomplished, and why it was deliberately abandoned — and what that sequence reveals about the company's decision-making.
The Fisheries Vertical Was a Deliberate Constraint, Not a Mistake
The fisheries monitoring business was never designed to be a large standalone company. The market for electronic monitoring software in commercial fisheries is structurally limited: the customer base is fragmented (thousands of small fishing operators, dozens of national regulatory bodies), procurement is slow (government and NGO cycles), and willingness to pay is constrained by the economics of the fishing industry itself.
OnDeck's advisors — Dr. Daniel Pauly, Dr. Rashid Sumaila, Dr. Amanda Vincent[8] — are world-class fisheries scientists, not enterprise software executives. The company's early funders — National Geographic, Schmidt Marine Technology Partners, the Government of Canada[11] — are mission-driven organizations, not growth-oriented venture investors. The $1.5M Ocean Supercluster grant[10] is a research and development instrument, not a commercial scaling vehicle.
The founders appear to have understood this. The fisheries vertical provided something more valuable than revenue: it provided a technically adversarial deployment environment that forced the team to solve the training data problem in a domain where no shortcuts existed. You cannot scrape labeled images of rare Pacific rockfish species from the internet. You cannot fine-tune a model on footage that doesn't exist. The constraint forced the VLM decomposition approach that became the company's core IP.
The $150K ARR Ceiling Was a Signal, Not a Failure
By January 2025, the fisheries business had reached approximately $150K ARR.[18] This figure is notable for what it implies about the ceiling: after nearly three years of operation, a $1.5M government grant, and deployments across three countries, the company had not broken $200K ARR in its original vertical.
The structural explanation is straightforward. Fisheries monitoring customers — governments, NGOs, Indigenous communities, commercial operators — are not high-velocity software buyers. The $1.5M Ocean Supercluster grant was awarded for a project with a total value of $3.5M including matching funds,[10] suggesting the company was operating in a grant-dependent rather than revenue-driven growth model. Grant funding and ARR are not the same thing, and the $150K ARR figure suggests the commercial revenue base was thin relative to the total capital deployed.
The pivot was not a response to crisis. The company had six-figure revenue in hand and a major government grant. It was a response to opportunity: the NeurIPS paper had demonstrated that the VLM-based approach generalized beyond fisheries, and the founders recognized that the technology had outgrown the vertical.[12] As Dungate noted about the NeurIPS results: "We've beaten the benchmarks for species identification, and we have a program deployed in other private industries."[25] The phrase "other private industries" is the tell — the pivot was already underway before it was announced.
The Co-Founder Departure Was an Early Stress Test
Matthew Leighton's departure in fall 2022 — less than six months after incorporation — is worth examining. Leighton returned to his PhD at SFU,[7] which suggests the departure was voluntary rather than conflict-driven. But a CFO departure at the pre-revenue stage removes financial discipline from the founding team at precisely the moment when capital allocation decisions matter most.
The company's subsequent funding strategy — heavy on grants and competition prizes rather than equity — may reflect this gap. CAD $170K+ in grants and awards by 2022,[26] followed by the $1.5M Ocean Supercluster grant in 2023, is a coherent strategy for a mission-driven environmental tech company. It is less coherent for a company that was already planning to pivot to enterprise AI infrastructure. The disposition of the Ocean Supercluster grant after the January 2025 pivot — whether it was retained, returned, or renegotiated — is not publicly disclosed, and this is a meaningful unknown.
The Structural Risk: Platform Commoditization
The most significant forward risk for OnDeck is not competition from other startups — it is the possibility that the major VLM providers build native video analysis capabilities that commoditize the core offering. OpenAI's GPT-4V and subsequent multimodal models already demonstrate qualitative visual reasoning. Google's Gemini 1.5 Pro can process hours of video natively. If these capabilities become accessible via standard API at commodity pricing, the "no training data required" differentiator narrows.
OnDeck's response to this risk — SOC 2 certification, air-gapped deployment, defense-grade security[19] — is the correct one. The major VLM providers are not going to offer air-gapped deployment to the Singaporean Military on a $160K pilot contract. The company's moat is not the model; it is the deployment infrastructure and the enterprise trust layer built around it. This is a defensible position, but it requires the company to move faster than the platform providers on enterprise compliance — a race that has a defined endpoint.
What the Fisheries Chapter Actually Accomplished
The fisheries vertical accomplished four things that the enterprise vision infrastructure business directly inherited: (1) it forced the team to solve the training data problem in a domain where no workaround existed; (2) it produced a peer-reviewed NeurIPS paper that validated the VLM approach independently of commercial claims; (3) it generated a differentiated founder narrative — real-world deployment on fishing vessels in Costa Rica is a more credible stress test than a lab benchmark; and (4) it attracted mission-driven funders (National Geographic, Eric Schmidt's foundation) who remain backers in the current phase.[22]
The fisheries chapter was not a detour. It was the R&D phase of a company that couldn't afford a traditional R&D phase.
Key Lessons
-
A niche vertical can function as a paid R&D environment if the technical constraint it imposes is generalizable. OnDeck's fisheries monitoring business forced the team to solve the training data problem in a domain where labeled datasets were structurally unavailable. The solution — VLM decomposition and recombination — proved applicable across defense, oil and gas, and robotics. A startup that had started in enterprise CV directly would have had access to richer training data and less pressure to innovate around the constraint. The adversarial conditions of the fisheries vertical produced the IP.
-
Grant funding and ARR measure different things, and conflating them delays the pivot decision. OnDeck received $1.5M from Canada's Ocean Supercluster in October 2023 while generating approximately $150K in commercial ARR — a ratio that suggests the business was grant-dependent rather than commercially self-sustaining. The pivot came in January 2025, roughly 15 months after the grant was awarded. Whether the grant accelerated or delayed the pivot is unknown, but the pattern — large non-dilutive funding masking thin commercial traction — is a common trap in mission-driven deep tech companies.
-
SOC 2 certification and air-gapped deployment are not compliance checkboxes — they are market access decisions. OnDeck's $160K Singaporean Military pilot would not have been possible without the security infrastructure the company built. Most AI vision startups treat enterprise compliance as a late-stage concern; OnDeck treated it as a prerequisite for the customer segments it was targeting. The Singaporean Military deployment is the direct result of that sequencing.
-
A NeurIPS workshop paper is a more durable asset than a press release. OnDeck's pivot narrative is credible in part because it is anchored to peer-reviewed research, not just founder claims. The paper demonstrating VLM superiority on niche species identification tasks gave the company an independent validation of its technical approach that enterprise customers and YC evaluators could assess without relying solely on the founders' word. In a market full of AI companies making unverifiable capability claims, third-party validation is a meaningful differentiator.
-
The "world's most powerful vision model" claim requires independent verification to be durable. OnDeck's Perception-0 is described on the company's website as "the world's most powerful vision model, guaranteed to maintain SOTA over time." This claim is currently unverified by independent benchmarks. As the company scales into defense and critical infrastructure contracts — where procurement teams conduct rigorous technical due diligence — the absence of third-party validation will become a friction point. The NeurIPS paper covers the fisheries-era methodology; the enterprise-era claims need equivalent external validation.
Sources
- Y Combinator — OnDeck AI company profile
- GeekWire — "Tracking fish: University students use machine learning for fishery monitoring startup" (June 14, 2022)
- VanTech Journal — Alexander Dungate / OnDeck Fisheries AI profile (April 14, 2025)
- GeekWire — "Fishery monitoring software wins $25K at Univ. of Washington student startup competition" (May 27, 2022)
- Medium — Melissa Bernstein, "My Experience Working as the First Software Engineer at OnDeck Fisheries AI" (February 1, 2024)
- New Ventures BC — "Top 25 Spotlight: OnDeck Fisheries AI" (August 4, 2022)
- BC Technology — "Vancouver-based OnDeck Fisheries AI Awarded $1.5 Million in Funding From Canada's Ocean Supercluster" (October 4, 2023)
- em4.fish — "Industry Focus: OnDeck Fisheries AI" (December 4, 2024)
- Creative Destruction Lab — OnDeck Fisheries AI Inc. profile (October 27, 2025)
- Y Combinator — OnDeck AI launch post (August 20, 2025)
- American Bazaar — "Y Combinator backs new wave of AI startups" (August 25, 2025)
- National Fisherman — Alexander Dungate quote on training data
- OnDeck AI website — Perception-0 and SOC 2 claims
- Work at a Startup — OnDeck AI Founding Engineer job listing
- Y Combinator — OnDeck AI jobs page
- PitchBook — OnDeck AI funding profile