Bringing Continual Learning to Every Company: Trajectory’s Ronak Malde

Core Philosophy & Market Gap

• Trajectory aims to bridge the gap between "plentiful intelligence" (foundation models) and "plentiful experience" (domain-specific human work).
• The goal is to shift the AI paradigm from a "pre-industrial" state, where everyone relies on a single generic oracle, to a state of "abundance" via specialization and continuous learning.
• The ultimate objective is to increase AI adoption from the current ~1% of enterprises to 99% by making model customization as simple as setting trellises for a self-growing plant.

Definition of Continual Learning

• Current State: Described as "pseudo continual learning," where updates require offline batch processing and consulting external services (analogous to visiting AAA for map updates).
• Goal State: "True continual learning," analogous to Google Maps, where systems update in real-time based on aggregate traffic and route data without manual intervention.
• Unit of Work: Proposes shifting the focus from "tokens" to "trajectories" (economically valuable tasks or rollouts in the real world).

Technical Architecture & Research

• Architectural Shifts: Acknowledges that base model architectures are evolving (e.g., DeepSeek v4.1 Flash) to include fast/slow memory and skip connections, which are prerequisites for continual learning.
• GRPO Algorithm Modification: Released research on a "group size one" paradigm for the GRPO algorithm.
  • Traditionally, GRPO requires replaying a task 8-16 times to establish a baseline; Trajectory's modification allows learning from a single rollout of 10,000 diverse trajectories.
  • Eliminates the need for complex, perfect environment simulations.
• Reward Modeling: Developed a "metagrader" approach.
  • Instead of labeling all traces, the system selects a representative subset (e.g., 100 out of 10,000).
  • The model learns a rubric from this small set and propagates the reward signal to the broader dataset.
• Curiosity & Alignment: Identifies "curiosity" (quantifying unknown transition functions in Markov Decision Processes) as a missing piece for autonomous agent learning.
  • Prioritizes "continual learning" (learning from human feedback loops) over "Recursive Self-Improvement" (RSI) to maintain alignment with human objectives.

Product Strategy & Interface

• Value Proposition: Democratizes access to frontier model training by consolidating training, evaluation, and deployment into a single interface.
• Key Primitives Required:
  1. Tracing: Logging full trajectories of production agents, sub-agents, and user feedback (often via partners like Blankchain, Braintrust, or Raindrop).
  2. Replayability: Converting traces into a reproducible environment (e.g., via Docker files) for evaluation.
  3. Algorithm Modification: Abstracting the complex training stack (loss functions, reward functions) which Trajectory handles internally.
• User Interface: Designed as a "one-click" system where users define traces, specify "what good looks like," and deploy updated models.
• Strategic Analogy: Compares the platform's approach to Apple combining the internet and phones; it simplifies the complex loop of data-to-intelligence for non-researchers.

Customer Success & Case Studies

• Harvey (Legal Tech):
  • Problem: High costs and inability to leverage 30 years of proprietary legal nuance using generic frontier models.
  • Result: Trained a custom model that is 6x cheaper, improving accuracy from ~4% (open source) to ~18%, with headroom for further growth.
• Rogo (Finance):
  • Focus: Optimizing for "intelligence density" and efficiency in autonomous finance tasks.
  • Result: Built models that are faster, cheaper, and more efficient in token usage while maintaining high precision.
• NVIDIA: Currently utilizing the platform for their agent infrastructure.

Organizational Structure & Team

• Team Composition: Unifies frontier researchers (ex-Gemini 3, GPT-4.0, DeepMind) with product engineers (ex-Stripe, Netflix).
• Productionized Research: Unlike traditional labs, Trajectory treats the training stack as a microservices-based Kubernetes system with auto-scaling, allowing design engineers to directly influence training runs.
• Workflow: Leverages frontier coding agents to handle implementation, freeing human engineers to focus on system design and whiteboarding.

Future Outlook

• Product Evolution: The product is expected to evolve from just model training to optimizing the entire "harness" (sub-agent routing and orchestration).
• Market Vision: Anticipates a future where companies route 90% of tasks to efficient, custom-trained models and only 10% to expensive frontier models.
• Ecosystem: Highlights the need for a supportive infrastructure ecosystem (e.g., Modal for sandboxes, Thinking Machines) to make continual learning viable.