Trajectory
by Trajectory
Continual-learning platform that post-trains open models on your production feedback and expert judgment
Trajectory is a continual-learning platform that post-trains open-source models on an application's own production traces, corrections and expert grades, so the model keeps improving after launch. It is for AI product teams in law, finance and support that want a specialised model without hiring a post-training research team.
Trajectory is a post-training platform built around continual learning, founded about a year before its November 2026 launch by Stanford classmates Ronak Malde (CEO, among the Windsurf researchers Google DeepMind acqui-hired), Michael Elabd (formerly DeepMind) and Arjun Karanam (formerly generative foundation models at Apple). Teams start from an open-source model, connect the platform to an existing agent harness with what the company says are two code changes, and stream production signals such as traces, user corrections and edits through an API or integrations like Zendesk. Domain experts define what good output looks like, Trajectory trains on those examples in one click, benchmarks the result against other leading models, and pushes an approved update to production. Every update passes through the customer's eval suite and approval workflow, and training runs are auditable. Research-preview features include Continual Grading, which applies a customer-defined grading standard across production traces, and Group Size 1, which trains on each production interaction. On November 2, 2026 the company opened its self-serve platform and announced a $45 million Series A led by Sequoia at a $300 million valuation, with NVIDIA, Bessemer Venture Partners and Bain Capital Ventures participating; its site cites $60 million raised across seed and Series A. TechCrunch named Harvey, Rogo, Decagon, Mercor and NVIDIA as customers, and the team was 18 people. It competes with Thinking Machines Lab's Tinker, Applied Compute and Prime Intellect.
An AI product or platform team running a production agent on an open model that has domain experts but no in-house reinforcement-learning researchers
A specialised model that is retrained on real production feedback, gated by your own evals, without building a post-training stack
At a Glance
- Category
- AI Models & APIs
- Pricing
- Contact for pricing
- Target Market
- CTOs, Heads of AI, ML Engineers, AI Product Teams
- Deployment
- Cloud-first, API-based
- Founded
- 2025
- Team Size
- 11-50
Key Features
- ✓Bring your own agent
Works with any existing agent harness and, per the vendor, needs two code changes to start collecting training signal.
- ✓Signals
Streams production traces, corrections and edits through an API or integrations such as Zendesk to use as training data.
- ✓One-click training
Turns expert-supplied examples of good output into a post-training run without writing reinforcement-learning code.
- ✓Compare
Benchmarks the tuned model against other leading models to show where it still falls short before deployment.
- ✓Gated deployment
Pushes model updates to production only after they pass the customer's eval suite and approval workflow, with auditable training runs.
- ✓Continual Grading (research preview)
Applies a grading standard the customer defines to production traces so the model keeps learning after launch.
Capabilities
Use Cases
- •Legal model specialisation
A legal AI vendor post-trains an open model on lawyer feedback so outputs match the firm's own standard of a good answer.
- •Financial research agent
A finance AI team trains on analyst corrections to improve a specialised model for research tasks, as Rogo does.
- •Support agent improvement loop
Ticket edits from Zendesk stream back as signals, and each retrained version ships only after passing the team's evals.
- •Open-model cost replacement
A team benchmarks a post-trained open model against frontier models and switches a narrow workload once quality gaps close.
Ideal For
Best For
- ✓Legal and financial AI products that need a model tuned to expert judgment, as Harvey and Rogo use it
- ✓Support agents that can feed ticket corrections from Zendesk back into training
- ✓Teams replacing an expensive frontier API with a post-trained open model for one narrow task
- ✓Organisations that want every model update to pass their own eval suite and approval step
Not Ideal For
- ✗Teams that rely only on closed frontier APIs, since the platform post-trains open-source models
- ✗Buyers who need published pricing, a list of supported base models or on-prem deployment today; none of these are documented and access still runs partly through a waitlist
- ✗Applications without meaningful production traffic or expert feedback to train on
Integrations
Deployment
Market Analysis
Pros
- ✓Named customers include Harvey, Rogo, Decagon, Mercor and NVIDIA
- ✓Removes the need for scarce post-training researchers by letting domain experts define good output
- ✓Model updates are auditable and pass customer evals before deployment
- ✓SOC 2 certified per the vendor's site
Cons
- ✗No published pricing, supported-model list or SDK language list in the docs
- ✗Site still routes new users to a waitlist, and production use where the app stays in your own infrastructure is listed as private beta
- ✗Very young company of about 18 people; Runtime Wire described Harvey, Rogo, Mercor and NVIDIA as testing the approach rather than confirmed paying customers
- ✗Hacker News discussion is minimal, so there is little independent production evidence yet
Pricing
Platform
Contact for pricing
- ✓Signals ingestion
- ✓One-click training
- ✓Model comparison
- ✓Gated deployment
No pricing is published on the site, in the docs or in launch coverage. The website routes new users to a waitlist even though TechCrunch called the platform publicly available on November 2, 2026, so expect a sales-led onboarding with costs quoted per engagement.
Security & Compliance
Connect
Sources
This page was written from 5 sources, 4 on domains other than trajectory.ai.
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
Heabsy
OpenAI- and Anthropic-compatible inference API for open-weight models, with an EEA-only tier
Liquid AI LFM3
Open-weight 2B text, vision, audio and decision models that run on phones, laptops, cars and private servers
Recursive Flint
Apache 2.0 open-weight coding model (40B MoE, 2.5B active) built to run coding agents locally on a laptop
Arena AI Evaluations
Paid model evaluations built on the Arena leaderboard's crowdsourced human votes, for enterprises choosing an LLM