Trajectory
A platform for continual learning, enabling AI models to improve from real-world user interactions.
Website: https://www.trajectory.ai/
Cover Block
Open sources
| Field | Value |
|---|---|
| Name | Trajectory |
| Tagline | A platform for continual learning, enabling AI models to improve from real-world user interactions [Trajectory.ai] |
| Headquarters | San Francisco, United States |
| Founded | 2026 |
| Stage | Series A [TechCrunch, November 2026] |
| Business model | SaaS |
| Industry | Deeptech |
| Technology | AI / Machine Learning |
| Geography | North America |
| Growth profile | Venture Scale |
| Founding team | Co-Founders (3+), Ronak Malde, Michael Elabd, Arjun Karanam [Wired, May 2026] |
| Funding label | $50M+ |
| Total disclosed funding | Approximately $60,000,000 [TechCrunch, November 2026] |
Links
Open sources
- Website: https://www.trajectory.ai/
What an Investor Needs First
PUBLIC Trajectory is building software that lets companies improve AI models from live product usage, and it merits attention now because it paired that thesis with a self-service launch in November 2026 and a $45 million Series A led by Sequoia announced the same day [TechCrunch, November 2026] [Trajectory.ai] [Introduction - Trajectory]. The company was founded in 2026 by Ronak Malde, Michael Elabd, and Arjun Karanam, Stanford classmates according to public profiles and reporting, with the early narrative centered on bringing research-grade continual learning into production workflows that nontechnical domain experts can use [Wired, May 2026] [Jerry Chan - LinkedIn, retrieved 2026] [TechCrunch, November 2026].
The product position is fairly specific: Trajectory says customers can start with an open-source model, connect it to existing applications, capture signals such as edits, corrections, and retries, and use that data to train, evaluate, and deploy better-performing models over time [Trajectory.ai] [Introduction - Trajectory] [Built In]. That framing matters because the stated wedge is not the base model itself, but the feedback loop and tooling around post-training, with company materials emphasizing compatibility with existing agent stacks and a low-integration workflow [Trajectory.ai] [TechCrunch, November 2026].
The founding team's relevance is stronger than average for a young infrastructure company. Wired reported that Malde previously worked at Windsurf and later joined Google DeepMind, Elabd worked in DeepMind's robotics division, and Karanam worked at Apple on generative foundation models and Vision Pro related efforts [Wired, May 2026].
On capitalization, public reporting points to a $15 million seed led by Conviction in May 2026 and a $45 million Series A led by Sequoia in November 2026, for roughly $60 million in disclosed funding to date, with SaaS the apparent business model and an 18-person team as of November 2, 2026 [Wired, May 2026] [TechCrunch, November 2026].
What matters over the next 12 to 18 months is less the fundraising pace than evidence that continual learning can become a repeatable product motion rather than a founder-led research service. Public references to partnerships with Clay, Harvey, Decagon, Mercor, and Rogo are promising, but investors will want clearer proof of production depth, renewal behavior, and whether nontechnical teams can in fact drive model improvement without heavy bespoke support [Yahoo Finance] [Pulse2.com] [Trajectory.ai].
Claim stands unchecked -- This section mixes independent reporting from TechCrunch and Wired with material product and partner claims that remain company-sourced or lightly corroborated in secondary coverage.
Taxonomy Snapshot
| Axis | Value |
|---|---|
| Stage | Series A |
| Business Model | SaaS |
| Industry / Vertical | Deeptech |
| Technology Type | AI / Machine Learning |
| Geography | North America |
| Growth Profile | Venture Scale |
| Founding Team | Co-Founders (3+) |
| Funding | $50M+ total disclosed, approximately $60,000,000 |
Inside the Company
PUBLIC
Trajectory emerged quickly in 2026 around a narrow but timely premise: model improvement should not stop at deployment. The company presents itself as a San Francisco-based research lab and product company building a platform for continual learning, with a focus on helping teams improve AI systems from real-world usage rather than one-off model releases [Trajectory.ai]. Structured company facts place the founding in 2026 and headquarters in San Francisco, United States [Trajectory.ai].
The public milestone record is short, but it is already legible. Wired reported in May 2026 that the company launched with a $15 million seed round led by Conviction, backing founders Ronak Malde, Michael Elabd, and Arjun Karanam, whose prior work spanned Windsurf, Google DeepMind, and Apple [Wired, May 2026]. By November 2, 2026, TechCrunch reported that Trajectory had launched a self-service platform and raised a $45 million Series A led by Sequoia, bringing disclosed financing to about $60 million [TechCrunch, November 2026]. On the company site, Trajectory describes the product in plain terms as "the platform for continual learning" and frames the workflow around connecting production systems, capturing feedback, and shipping model improvements over time [Trajectory.ai].
Partially corroborated -- Company description is supported by the company website, while founding chronology and financing milestones in this section rely primarily on company materials plus single-publication reporting from Wired and TechCrunch.
Under the Hood
Reported and inferred
Trajectory is making a specific bet on where AI application value accrues: not at initial model selection, but in the feedback loop after deployment. Public materials describe the product as a continual-learning platform that lets teams connect an open-source model to an existing application, capture production signals such as edits, corrections, retries, and traces, and use that data to improve model behavior over time [Trajectory.ai] [TechCrunch, November 2026] [Built In]. The company launched a self-service version of that platform on November 2, 2026, according to TechCrunch, which matters because it shifts the product from an assisted research workflow toward software that customers can adopt directly [TechCrunch, November 2026].
The product surface described across the website and docs is relatively consistent. Trajectory says its SDK converts raw agent traces and product telemetry into a standardized format used for training, evaluation, and continual learning, while the website shows modules for signals, datasets, models, training runs, evaluations, graders, and deployments [Introduction - Trajectory] [Trajectory.ai]. The same materials say the system works with any agent harness and requires two code changes to get started, though that ease-of-integration claim remains company-sourced and should be read as positioning rather than independently verified implementation evidence [Trajectory.ai].
The technical differentiation, as presented publicly, is less about building a frontier base model and more about operationalizing post-training for domain teams. Company and press descriptions repeatedly frame the platform as a way for nontechnical experts to shape model behavior using production feedback, reducing the need for in-house research talent or forward-deployed ML engineering in specialized settings such as legal, customer service, and go-to-market workflows [Trajectory.ai] [TechCrunch, November 2026] [Zetik] [Built In]. Ronak Malde has also publicly discussed on-policy self-distillation, or OPSD, as part of the company's view of continual learning, including token-level feedback and lower infrastructure overhead, but those comments come from interviews and secondary summaries rather than a peer-reviewed technical release tied directly to the product in production [StartupHub.ai, 2026] [BigGo Finance] [SozAI].
Partially corroborated -- Product functionality is supported by Trajectory's website and documentation, with partial corroboration from TechCrunch and Built In. The more technical claims around OPSD and implementation ease rely mainly on company statements and interview coverage.
Market Research
Open sources
The market matters now because the cost of putting AI into production has shifted from model access to model adaptation, and Trajectory is aimed at that second layer rather than the foundation model itself [TechCrunch, November 2026] [Wired, May 2026].
There is no confirmed third-party TAM, SAM, or SOM figure in the sourced record for continual-learning platforms specifically, so the cleaner approach is to frame Trajectory against adjacent public markets rather than force precision that the evidence does not support. The public reporting consistently places the company in the workflow of post-training open-source models on live usage data, with an emphasis on making that process usable by domain experts instead of dedicated ML researchers [TechCrunch, November 2026] [Trajectory.ai]. That positions the company at the intersection of several analogous markets: enterprise AI software, model training and evaluation tooling, and the emerging software layer around AI agents and production feedback loops [TechCrunch, November 2026] [Built In].
Demand drivers are visible even without a formal market-size study in the source set. Multiple sources describe the same operational problem: specialized AI applications often need repeated tuning after deployment, because benchmark gains alone do not capture how users correct, retry, or edit model outputs in real workflows [Zetik] [Built In]. Trajectory's pitch, according to public reporting and its own product materials, is that those production signals can become training data rather than remaining exhaust, which is a meaningful proposition for teams in legal, customer service, and go-to-market use cases where preferences shift by company and by task [TechCrunch, November 2026] [Built In] [Trajectory.ai].
That framing also clarifies the adjacent and substitute markets. One substitute is the status quo of hiring in-house research talent or forward-deployed ML engineers to fine-tune models for each vertical application, a bottleneck identified in third-party coverage of the company [Zetik]. Another is relying on major foundation-model releases to lift application quality in steps rather than building a continual post-training loop from product telemetry, which is the contrast Trajectory's CEO has made in external interviews about "continual learning" and on-policy self-distillation [BigGo Finance] [StartupHub.ai, 2026]. In practice, the company is not competing only with software vendors. It is also competing with internal tooling, bespoke services work, and the option to defer optimization altogether.
Macro and regulatory forces cut both ways. The tailwind is that more enterprises are moving AI systems into user-facing workflows, which increases the volume of traces, edits, corrections, and evaluations that can feed back into model improvement [Trajectory.ai] [Built In]. The constraint is that any product built on production data will face procurement scrutiny around data handling, auditability, and model-governance expectations, especially in regulated or high-stakes domains such as legal and potentially healthcare or finance, which are cited as relevant application areas in coverage of the category problem [Zetik]. The available sources do not establish Trajectory's compliance posture, so that remains a market access question rather than a disqualifier.
| Market lens | What the sources support | Source basis |
|---|---|---|
| Continual learning software | Product category framed around post-training models on real-world usage signals | [TechCrunch, November 2026] [Trajectory.ai] |
| Model training and evaluation tooling | SDK standardizes traces and telemetry for training, evaluation, and continual learning | [Introduction - Trajectory] |
| Vertical AI optimization layer | Targets teams in customer service, go-to-market, and legal workflows where behavior is domain-specific | [Built In] [Zetik] |
| Services substitute | Alternative remains hiring internal researchers or forward-deployed engineers for specialization work | [Zetik] |
The practical read is that Trajectory appears to be pursuing a budget line that may emerge from several adjacent software buckets rather than a single clean market category. That can support demand if buyers already feel the pain of post-deployment model drift, but it also means category education is likely part of the go-to-market burden.
Claim stands unchecked -- Market sizing is not supported by third-party quantified reports in the source set, and much of the section rests on company materials plus interpretive third-party coverage from TechCrunch, Built In, and Zetik.
Competition and Substitutes
Positioning
MIXED Trajectory appears to sit in an emerging layer between foundation model vendors and application teams, selling the tooling and workflow for continual post-training from production feedback rather than the base model itself [TechCrunch, November 2026] [Trajectory.ai] [Introduction - Trajectory].
The immediate competitive map is easier to define by category than by named company, because the source set does not identify direct rivals by name. At one end are foundation model providers, which improve general-purpose models on their own release cycles and can absorb parts of the post-training stack into their platforms over time. At the other end are internal ML teams and forward-deployed research or engineering functions, which remain the default substitute when a company wants domain-specific model improvement but does not buy a dedicated platform [TechCrunch, November 2026] [Zetik]. In between sit application-layer AI companies such as Clay, Harvey, Decagon, Mercor, and Rogo, which sources describe as partners or design partners rather than competitors, and which matter competitively because some could choose to build similar feedback-to-training infrastructure in-house if the economics or control requirements justify it [Yahoo Finance] [Pulse2.com] [Dragdropship].
That framing matters because Trajectory's wedge is not model ownership. The company's public pitch is that businesses can start from open-source models, connect those models to existing applications, capture signals such as edits, retries, corrections, and traces, then use that data to improve performance over time without relying entirely on specialized ML researchers [Trajectory.ai] [Introduction - Trajectory] [Built In] [TechCrunch, November 2026]. If that claim holds in practice, Trajectory competes less with a single startup and more with a bundle of alternatives: buying more from a foundation model vendor, staffing an internal post-training team, or accepting slower model improvement cycles.
The defensible edge today appears to be a combination of founder credibility, timing, and early access to production feedback loops from AI-native partners. Wired identified the founding team as Ronak Malde, Michael Elabd, and Arjun Karanam, with prior experience spanning Windsurf, Google DeepMind, and Apple, a background that is relevant for a technically demanding product category where buyers are likely to scrutinize research depth before entrusting training pipelines [Wired, May 2026]. Public sources also say Trajectory is working with companies including Clay, Harvey, Decagon, Mercor, and Rogo, which, if active and sticky, could give the platform exposure to live domain-specific data exhaust that is difficult for a new entrant to replicate quickly [Yahoo Finance] [Pulse2.com] [Daytona]. The durability of that edge is less certain. Talent and early design-partner access are meaningful at this stage, but both are perishable if larger model platforms ship adjacent tooling, or if customers decide their production feedback loops are strategic enough to keep in-house.
The sharpest exposure is not a named startup in the current source set, but the structural advantage held by the layers above and below Trajectory. Foundation model vendors own distribution, model weights, and often the default developer relationship; internal platform teams at large AI application companies own proprietary workflow data and can tailor systems to their own latency, privacy, and cost constraints. Trajectory does not appear, from public reporting, to own an end-customer application channel or a proprietary foundation model, so it needs to keep proving that a neutral continual-learning layer is easier and faster to adopt than building internally or expanding spend with a model vendor [TechCrunch, November 2026] [Trajectory.ai]. That also implies some categories may remain hard to penetrate, especially where data governance, model sovereignty, or highly customized infra requirements push buyers toward internal builds.
Over the next 18 months, the most plausible competitive scenario is a separation between workflow owners and infrastructure specialists. Winner if X: Trajectory, if AI-native software companies continue to prefer open-source model stacks and need a dedicated system for converting production behavior into repeatable post-training gains faster than internal teams can build it [Trajectory.ai] [TechCrunch, November 2026]. Loser if Y: the in-house substitute at smaller AI application companies, if a self-service product launched in November 2026 is good enough to displace early internal tooling and reduce the need for dedicated research hires [TechCrunch, November 2026]. The reverse scenario is also credible. If foundation model vendors fold similar capabilities into their managed platforms, or if design partners conclude that feedback data is too strategic to externalize, Trajectory's position could narrow from category layer to feature supplier.
Opportunity
Upside case
PUBLIC The prize here is not another model tool, but the chance to become the control layer that tells specialized AI systems how to get better after deployment, across legal, customer service, go-to-market, and other workflow-heavy categories where production feedback is both frequent and expensive to waste [TechCrunch, November 2026] [Built In] [Zetik].
The headline opportunity is straightforward. If Trajectory can make continual post-training as usable as analytics or observability software, it could become default infrastructure for companies that run AI products on top of open-source models and need those systems to improve from day-to-day usage rather than from occasional frontier-model refreshes [TechCrunch, November 2026] [Trajectory.ai] [Introduction - Trajectory]. That outcome is reachable, not merely aspirational, because the company already has three ingredients young infrastructure vendors usually need before category leadership is even discussable: a product in market through its self-service launch on November 2, 2026, founders with direct model-training experience from DeepMind, Windsurf, and Apple, and a reported set of AI-native partners including Clay, Harvey, Decagon, Mercor, and Rogo that maps onto real production use cases rather than lab demos [TechCrunch, November 2026] [Wired, May 2026] [Yahoo Finance] [Pulse2.com].
The most important question is how that upside branches from here. In practice, there are a few credible paths to very large scale, and each depends on a different catalyst.
| Scenario | What happens | Catalyst | Why it's plausible |
|---|---|---|---|
| Workflow system of record for AI improvement | Trajectory becomes the standard layer where teams collect traces, corrections, edits, and evaluations, then turn them into repeated model updates across production applications | The November 2026 self-service launch lowers adoption friction from bespoke engagements to product-led deployment [TechCrunch, November 2026] | The company says its SDK standardizes raw agent traces and telemetry for training and evaluation, and the product claims compatibility with any agent harness with two code changes, which is the kind of integration story that can widen top-of-funnel if it works in practice [Introduction - Trajectory] [Trajectory.ai] |
| Embedded engine for AI-native leaders | A handful of fast-growing AI application companies use Trajectory deeply enough that it becomes part of their core model-improvement stack, then reference adoption pulls in adjacent customers | Design-partner or production validation from named AI-native companies such as Clay, Harvey, Decagon, Mercor, and Rogo [Yahoo Finance] [Pulse2.com] [Dragdropship] | Public reporting already places Trajectory alongside customers or partners in specialized domains where user feedback is abundant and model behavior matters commercially, which is where continual learning should have the clearest return on effort [Built In] [Zetik] [Daytona] |
| Category standard for nontechnical post-training | Trajectory expands the buyer from ML researchers to domain teams, making legal, support, and go-to-market operators active participants in improving models | The company convinces customers that domain experts can guide post-training without building internal research teams [TechCrunch, November 2026] [Trajectory.ai] | Both the company materials and TechCrunch coverage anchor on accessibility for nontechnical experts, and third-party descriptions repeatedly frame the bottleneck as reliance on in-house researchers or forward-deployed engineers [TechCrunch, November 2026] [Trajectory.ai] [Zetik] |
The compounding mechanism is plausible because the product sits in the feedback loop itself. Each deployment can generate structured traces, corrections, edits, retries, and evaluations that make future training runs more useful, while the customer's own experts get better at specifying what good behavior looks like over time [Trajectory.ai] [Introduction - Trajectory] [Built In]. If Trajectory becomes the place where those signals are captured, normalized, scored, and turned into repeatable post-training workflows, switching costs could grow from simple API integration to accumulated operating knowledge: datasets, graders, evaluation histories, and deployment patterns all living in one system [Trajectory.ai] [Introduction - Trajectory]. The early signs are modest but real, namely a self-service product now in market, reported production-oriented partnerships, and messaging that consistently centers on observed product usage rather than synthetic benchmark improvement [TechCrunch, November 2026] [Yahoo Finance] [BigGo Finance].
The size of the win is easiest to frame through infrastructure economics rather than a market-size figure, because no confirmed TAM source is in the public record provided here. The company has raised about $60 million across its seed and Series A, including a reported $45 million Series A at a $300 million valuation led by Sequoia, which places it in the cohort of venture-backed AI infrastructure bets expected to chase outsized platform outcomes rather than narrow tooling exits [TechCrunch, November 2026] [Wired, May 2026]. If the strongest scenario plays out, where Trajectory becomes a standard continual-learning layer for AI-native and later enterprise applications, the upside could be measured in multibillion-dollar enterprise value terms (scenario, not a forecast), because core model-improvement infrastructure tends to command platform-like valuations when it becomes embedded in production workflows and trusted by domain operators. The evidence today supports the possibility of that path, but not yet its inevitability.
Partially corroborated -- This section relies on one strong independent report from TechCrunch, one earlier independent report from Wired, and several company or lightly corroborated third-party descriptions for product behavior and partner references [TechCrunch, November 2026] [Wired, May 2026] [Trajectory.ai] [Yahoo Finance].
Sources
Open sources
[Trajectory.ai] Trajectory | https://www.trajectory.ai/
[TechCrunch, November 2026] Trajectory raises $45M to make AI model training accessible to nontechnical experts | https://techcrunch.com/2026/11/02/trajectory-raises-45m-to-make-ai-model-training-accessible-to-nontechnical-experts/
[Introduction - Trajectory] Introduction - Trajectory | https://docs.trajectory.ai/introduction
[Wired, May 2026] Former Google and Apple Researchers Launch a Startup ... | https://www.wired.com/story/ex-google-apple-ai-researchers-want-to-make-ai-that-gets-smarter-as-you-use-it/
[Jerry Chan - LinkedIn, retrieved 2026] Jerry Chan | https://www.linkedin.com/
[Built In] Built In | https://builtin.com/
[Yahoo Finance, retrieved 2026] Yahoo Finance | https://finance.yahoo.com/
[Pulse2.com] Pulse 2.0 | https://pulse2.com/
[Zetik] Zetik | https://zetik.com/
[StartupHub.ai, 2026] Trajectory's Ronak Malde on Scaling Continual Learn… | https://www.startuphub.ai/ai-news/ai-research/2026/trajectory-s-ronak-malde-on-scaling-continual-learning
[BigGo Finance] Scaling up Continual Learning | https://finance.biggo.com/podcast/dbf6d45da8ffc912
[SozAI] Scaling up Continual Learning | https://sozai.app/transcript/scaling-continual-learning-ronak-malde-trajectory/
[Dragdropship] Dragdropship | https://dragdropship.com/
[Daytona] Daytona | https://www.daytona.io/
Articles about Trajectory
- Trajectory's $60 Million Bet Turns User Corrections Into AI Model Updates — The ex-DeepMind team raised from Sequoia and Nvidia to make model retraining accessible to lawyers and analysts, not just ML engineers.