Trajectory has launched a self-service platform designed to help businesses train and improve AI models using feedback from everyday user interactions. The startup’s approach turns corrections, edits, retries and failed AI responses into training signals that companies can use to make their models better at specialised tasks. The platform aims to make AI model post-training accessible to business teams without requiring them to hire large teams of machine-learning researchers.
The launch, announced on November 2, 2026, came alongside a $45 million Series A funding round led by Sequoia Capital, valuing the company at $300 million. The round also attracted participation from Nvidia, Bessemer Venture Partners and Bain Capital Ventures. Trajectory wants to help companies move beyond static AI models towards systems that improve through experience and learn the specific requirements of the organisations using them.
Key Highlights
- Trajectory launched its self-service AI post-training platform on November 2, 2026.
- The startup raised $45 million in a Series A funding round led by Sequoia.
- The funding round valued Trajectory at $300 million.
- Nvidia, Bessemer Venture Partners and Bain Capital Ventures also participated.
- The platform collects feedback signals such as edits, corrections, retries and user interactions.
- Businesses can use these signals to improve AI models for specialised tasks.
- Customers include Harvey, Rogo, Decagon, Mercor and Clay.
- The company is targeting businesses that want to improve AI performance without building large internal research teams.
What Is Trajectory’s New AI Training Platform?
Trajectory is an AI infrastructure startup founded by Ronak Malde, Arjun Karanam and Michael Elabd, who have backgrounds at companies including Google DeepMind and Apple.
The company’s central idea is that general-purpose AI models are often capable of performing many tasks but do not automatically learn from mistakes made during everyday use.
For example, an AI assistant working for a law firm might repeatedly misunderstand how a particular organisation reviews contracts. A financial AI agent might also produce valuations that do not follow its customer’s preferred methodology.
Traditional solutions involve hiring machine-learning specialists to collect examples, analyse model performance and retrain the system. This process can be expensive and time-consuming.
Trajectory wants to simplify that process through a platform that captures real-world feedback and uses it to improve models.
Its self-service platform allows businesses to connect their existing AI applications, collect performance information and run post-training workflows without needing to build the entire training infrastructure themselves.
How Does Trajectory Turn Everyday Feedback Into AI Training?
The platform works by collecting information about how users interact with AI-generated results.
This includes several types of feedback:
- Edits: Changes users make to AI-generated responses.
- Corrections: Information indicating that an AI-generated result was incorrect.
- Retries: Instances where users ask the AI to attempt a task again.
- Acceptance: Cases where users approve or use the generated output.
- Task traces: Records of the steps an AI agent takes to complete a task.
These signals help companies understand where their AI systems perform well and where they need improvement.
Trajectory then uses the collected information to support model post-training.
The process can be understood in four stages:
- Connect the AI application: Businesses integrate Trajectory using its software development kit (SDK).
- Collect feedback: The platform captures relevant interactions, edits and corrections.
- Evaluate performance: Companies define what successful output looks like and assess model performance against those requirements.
- Train and deploy improvements: The system uses the feedback to refine models, with evaluation and approval before updates reach production.
The company also promotes a feature called Continual Grading, which aims to convert human preferences and expert judgments into repeatable evaluation signals.
This is particularly important in industries where there is no single correct answer. A legal document review, for example, may need to follow a particular firm’s interpretation, risk tolerance and internal standards.
Why Is Continual Learning Important for AI?
Most general-purpose AI models are trained on large datasets before being made available to users.
Once deployed, their underlying model weights generally remain unchanged until the developer releases an updated version or performs additional training.
This creates a limitation for businesses.
An AI assistant may learn the instructions given during a conversation, but that does not necessarily mean it will remember and apply the correction in future sessions.
Continual learning attempts to address this problem by creating a feedback loop between real-world usage and model improvement.
Instead of waiting for a completely new foundation model, companies can use their own operational data to refine AI systems for specific tasks.
For example, a customer support company could identify conversations in which an AI agent failed to resolve a refund request. Those interactions could become training examples that help improve future customer support responses.
Trajectory’s approach is designed to make this process repeatable, with post-training cycles potentially running weekly or more frequently.
However, this should not be confused with models changing their underlying weights instantly after every user interaction. Training and deployment still involve evaluation cycles and approval procedures.
Harvey, Rogo and Decagon Are Already Using Trajectory
Trajectory has attracted customers across several industries, including legal technology, financial research, sales and customer service.
Some of its customers include:
| Company | Business area | Potential application |
|---|---|---|
| Harvey | Legal AI | Improving legal research and contract analysis |
| Rogo | Financial AI | Refining financial research and analysis |
| Decagon | Customer service AI | Improving automated customer support |
| Mercor | AI talent and evaluation | Specialised AI workflows |
| Clay | Sales and data enrichment | Improving AI-powered business workflows |
| Nvidia | AI infrastructure and technology | AI model development and post-training |
Harvey and Rogo are among the companies using Trajectory to post-train open-source models for specialised tasks.
For example, a legal AI platform could use previous contract reviews to help a model better understand a firm’s preferred approach to particular clauses.
Similarly, a financial AI assistant could learn from historical company valuations and analyst feedback to improve how it performs specialised research.
The goal is to make these systems more useful for individual organisations rather than relying entirely on the capabilities of a general-purpose model.
Trajectory Raises $45 Million at a $300 Million Valuation
The platform launch coincided with Trajectory announcing a $45 million Series A funding round.
Sequoia led the round, with participation from Nvidia, Bessemer Venture Partners and Bain Capital Ventures.
| Funding detail | Information |
|---|---|
| Latest funding round | Series A |
| Amount raised | $45 million |
| Lead investor | Sequoia Capital |
| Other investors | Nvidia, Bessemer Venture Partners, Bain Capital Ventures |
| Reported valuation | $300 million |
| Founders | Ronak Malde, Arjun Karanam, Michael Elabd |
| Company focus | AI post-training and continual learning |
The company had previously raised a $15 million seed round at a reported $115 million valuation.
The latest funding provides resources to expand its engineering and research capabilities, develop its training infrastructure and make the platform available to a wider range of businesses.
The participation of Nvidia is also notable because the chipmaker has a major position in the infrastructure powering AI model development and deployment.
Trajectory Faces Competition From Other AI Training Startups
Trajectory is entering a growing market focused on making AI models easier to customise and improve.
One of its competitors is Thinking Machines Lab, founded by former OpenAI executive Mira Murati. The company’s Tinker platform provides infrastructure for developers who want to fine-tune and train AI models.
Other companies operating in related areas include Applied Compute and Prime Intellect.
These companies are targeting different parts of the AI development and training process, but share a broader objective: making advanced model development more accessible to businesses and developers.
Trajectory’s differentiation is its emphasis on feedback from production applications.
Rather than relying only on manually prepared training datasets, the company wants to capture signals generated when real users interact with AI products.
This could create a continuous source of task-specific training information, particularly for enterprise applications.
Data Privacy and Training Quality Remain Important Challenges
Trajectory’s approach also introduces important questions about data governance.
Production interactions can contain confidential business information, personal data, financial records and sensitive customer conversations.
Companies must therefore decide which information can be used for training and ensure that appropriate permissions and safeguards are in place.
Trajectory says its platform is SOC 2 certified and provides controls over which data is used for training. It also says model updates go through evaluation and approval workflows before being deployed.
Another challenge is the quality of user feedback.
Not every correction is necessarily accurate, and user preferences may differ across teams or customers. A model trained on incorrect feedback could become less reliable rather than more effective.
This makes evaluation systems and human oversight important parts of the training process.
The Bigger Picture
Trajectory’s launch reflects a shift in the AI industry from building increasingly powerful general-purpose models towards improving how those models perform in specific business environments.
Companies increasingly want AI systems that understand their internal processes, follow their preferred standards and improve as they handle more tasks.
If platforms such as Trajectory make post-training easier and less expensive, smaller businesses could gain access to capabilities that previously required specialised research teams.
The opportunity is substantial, but success will depend on whether companies can demonstrate measurable improvements in accuracy, cost and productivity without compromising privacy or reliability.
Looking Ahead
Trajectory will need to demonstrate that its self-service platform can consistently improve AI performance across different industries. Its ability to convert everyday feedback into useful training data, while maintaining strong evaluation and security controls, will be central to attracting more enterprise customers. The new funding should help the company expand its technology and compete with other AI training infrastructure providers.
For the broader AI industry, the launch highlights an important development: the next stage of competition may involve not only building larger models but also making existing models better at the specific jobs they perform. If continual learning becomes easier to implement, businesses could increasingly develop AI systems that improve through real-world experience rather than depending exclusively on periodic releases from major AI laboratories.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
