Trajectory Raises $45M Series A to Democratize AI Training

Training an artificial intelligence model to understand specialized professional work has historically required hiring machine learning scientists who charge six-figure consulting fees. When a medical hospital network wants an algorithm to identify rare tumor scans, or when a corporate legal firm needs software to parse offshore tax shelters, standard general-purpose models fail. Foundation models trained on public internet text lack deep institutional knowledge. Adapting those models to specialized professions previously required months of manual reinforcement learning, dataset curation, and continuous model retraining by specialized programmers. Silicon Valley startup Trajectory arrived on Monday, November 2, 2026, to automate that specialized training pipeline. The one-year-old venture announced that it raised $45M in a Series A financing round led by Sequoia Capital, valuing the eighteen-person enterprise at $300M. Strategic participation came from semiconductor giant Nvidia, alongside prominent institutional venture firms Bessemer Venture Partners and Bain Capital Ventures. Trajectory built a self-serve continual-learning platform that allows non-technical domain experts: including practicing physicians, courtroom attorneys, and tax accountants: to fine-tune foundation models directly through everyday work feedback. The funding round arrives as enterprises seek tailored machine learning solutions, continuing early-stage software financing trends we followed when Numeral secured $100M for corporate tax automation.

The investment represents an aggressive bet that automated post-training will replace expensive machine learning consulting contracts. Founded by former Google DeepMind and Apple systems engineers Ronak Malde, Michael Elabd, and Arjun Karanam, Trajectory addresses the primary bottleneck slowing down enterprise software adoption: the cost of keeping models accurate over time. In traditional machine learning operations, once a model is deployed, its performance degrades as real-world conditions shift. Retraining the model requires taking it offline, cleaning fresh data, and paying machine learning engineers to run expensive training runs on cloud server clusters. Trajectory platform automates this loop: connecting open-source models directly to enterprise production environments, collecting feedback from domain professionals as they edit documents, and retraining the model continuously in the background. Early enterprise adopters: including legal technology leader Harvey, financial analysis platform Rogo, customer service developer Decagon, hiring platform Mercor, and Nvidia itself: have deployed Trajectory to keep specialized models accurate with zero developer intervention. Specialized software automation across corporate administrative tasks matches models we examined when Thally deployed automated knowledge layers across enterprise software suites.

The Consulting Bottleneck: Why Enterprise AI Adoption Stalls

To grasp why Sequoia Capital committed forty-five million dollars to an eighteen-person startup, one must examine why corporate machine learning projects routinely fail. Over the past twenty-four months, enterprise organizations spent billions purchasing access to foundation models.

Chief information officers discovered that off-the-shelf models produce generic, shallow outputs when applied to specialized commercial tasks. A foundation model cannot draft an audit-compliant corporate merger contract or interpret a specialized hospital oncology record without extensive customization. To bridge that gap, enterprises hired external machine learning consultancies, paying hundreds of thousands of dollars to build custom fine-tuning pipelines. These consulting arrangements are slow and fragile. When an insurance regulation changes or a new pharmaceutical drug is approved, the custom model becomes outdated, requiring another round of consulting contracts to update training data. Trajectory eliminates the consultant middleman. The software integrates directly into the daily software tools used by professionals: such as electronic health record systems or legal document editors. When a doctor corrects an automated clinical summary, or an attorney revises a contract clause, Trajectory captures the correction as training feedback, automatically updating the model weights overnight. Continual learning turns daily workplace corrections into automated training data. Labor productivity adjustments across corporate enterprises were analyzed in our review of corporate technology deployments replacing administrative hours.

The Continual-Learning Architecture: Retraining on Production Data

The technical core separating Trajectory from legacy machine learning operations platforms sits in its continual-learning engine. Traditional fine-tuning frameworks like LoRA or full-parameter retraining require batch processing: collecting thousands of labeled examples, formatting them into training files, and running compute-heavy backpropagation passes on server clusters.

Trajectory engineered an automated pipeline that operates continuously through three functional stages:

• Passive Feedback Telemetry: The software monitors how domain professionals interact with model outputs, recording accepted edits, rejected suggestions, and manual corrections without requiring manual data labeling.
• Automated Data Quality Filtering: Neural filters evaluate incoming user feedback, discarding noisy corrections, contradictory inputs, and malicious prompt injections to ensure that only verified professional knowledge reaches the training queue.
• Incremental Parameter Optimization: The platform applies low-rank parameter updates to open-source foundation models during off-peak cloud hours, updating model weights without causing catastrophic forgetting of baseline general reasoning skills.

Because the training loop operates automatically, the model becomes more accurate the more it is used by domain staff. A hospital network that deploys Trajectory sees its clinical documentation accuracy improve week after week, as the system learns the specific terminology, treatment preferences, and formatting standards favored by local medical staff. Enterprise knowledge layers and automated software tools were explored in our analysis of CoComply automating contract compliance audits.

Comparing Enterprise Model Fine-Tuning Approaches