Trajectory, an 18-person startup that retrains open-source AI models on feedback from the applications they run in, has raised $45 million in a Series A round at a $300 million valuation, led by Sequoia with Nvidia, Bessemer Venture Partners and Bain Capital Ventures, TechCrunch reported on 2 November. In its own August field report, Nvidia’s Nemotron 3.5 Lightning, post-trained on Trajectory’s platform, passed 8.3% of held-out tasks on Harvey’s Legal Agent Bench, against 6.6% for Anthropic’s Claude Opus 4.6.
That gap is about two tasks, by our arithmetic. The report says it measured five tasks per category across 24 practice areas, and that the post-trained model produced ten passes: 10 of 120 tasks is 8.3%, and 6.6% of 120 is about eight. Taking the $300 million as post-money, as The Information reported it in August, it is 2.6 times the $115 million post-money valuation The Information reported for Trajectory’s $15 million seed round in the spring, according to Dealroom’s summary, also by our arithmetic.
Businesses using Trajectory start with an open-source model and connect the platform to their applications, which collect feedback on how the model performs; Trajectory then uses that feedback to retrain it, according to TechCrunch. Its homepage says “Trajectory works with any agent harness and takes just two code changes to get started.”
TechCrunch says companies adapting models for law or finance typically fine-tune general models, which often means hiring internal research teams or consultants known as forward-deployed engineers. Trajectory wants to automate that work. Co-founder Ronak Malde told TechCrunch that the people using it decide “what good looks like for their organization and can turn that into improvements for their model”.
Malde compared foundation models to a very smart PhD student who keeps starting a new job: “But what the economy values is someone who worked on the job for 30 years”, he told TechCrunch. Harvey and Rogo use the platform to post-train open-source models, and its other customers include Decagon, Mercor and Nvidia, according to the report, which says the platform is now publicly available.
All-pass rate on held-out Harvey Legal Agent Bench tasks (share of tasks meeting every rubric criterion), base model and after post-training on the Trajectory platform, as reported in Trajectory’s field reports of 11 August and 9 June 2026; Sonnet 4.6 and Opus 4.6 are closed models, not post-trained. Company-reported, not independently verified. The June report gave the Nemotron 3 Ultra base as 0%.
Business Pill · Fine-Tuning Explained
An evergreen Business Engineer explainer on fine-tuning, adapting a general AI model to one job. It teaches the general idea only and does not discuss Trajectory or its results.
The key insight: By our arithmetic, the result behind Trajectory’s pitch is about two tasks: its post-trained Nemotron 3.5 Lightning passed 8.3% of held-out legal tasks, ten of 120, against 6.6% for Claude Opus 4.6, about eight, according to Trajectory’s own August report; on a post-money basis, the round values the company at 2.6 times its seed valuation as The Information reported it.
The Round Against the Earlier Reports
In August, The Information reported a $40 million round at a $300 million post-money valuation, led by Sequoia with Nvidia and Bessemer, according to Dealroom’s summary of that report, which put the company at 12 employees. The round TechCrunch reported on 2 November is $45 million at a $300 million valuation, also names Bain Capital Ventures, and gives the headcount as 18. TechCrunch does not say whether $300 million is a pre-money or post-money figure.
The seed round, announced in late May, was $15 million with Conviction, Bessemer and Radical Ventures, plus Jeff Dean and Fei-Fei Li, according to Pulse 2.0; The Information reported its post-money valuation as $115 million, according to Dealroom. Disclosed funding comes to $60 million, by our arithmetic. Applied Compute, which TechCrunch names as a competitor alongside Thinking Machines Lab and Prime Intellect, was valued at $3.25 billion last month, about 11 times Trajectory’s valuation by our arithmetic.
What the Legal Benchmark Shows
Trajectory has published three field reports on post-training Nvidia’s open Nemotron models on Harvey’s Legal Agent Bench, an open benchmark of 1,200 agent tasks across 24 practice areas. The headline measure is the all-pass rate, the share of held-out tasks where an agent satisfies every rubric criterion.
The August report says “Post-trained Nemotron 3.5 Lightning reaches 8.3%, up from 0% at baseline”. It gives Nemotron 3 Ultra as moving from 0.8% to 5.8% and Nemotron 3 Super from 0.8% to 3.3%. The June report on Ultra gave its base as 0% and put Claude Sonnet 4.6 at 4.2% and Opus 4.6 at 6.6%.
On the share of individual rubric criteria passed, the June report puts post-trained Ultra at 83%, against 85% for Sonnet 4.6 and Opus 4.6 and 78% for GPT 5.5. It says Harvey’s post-trained Nemotron models matched the closed models’ quality “while costing at least 10x cheaper to run”, plotted against price per token.
The August report also says “Nine of 24 areas improve; none regress.” and adds a caveat: “At five tasks per category the resolution is coarse”. On output length, it says the base model used 22,000 tokens per held-out task and passed none, and that “the first post-trained model converts 90k tokens into ten passes”; a second run held the same pass rate at 37,000 tokens.
On 2 November, Palantir’s results presentation charted a fine-tuned 4-billion-parameter Nemotron at about 73% at sorting a construction firm’s requests for information, against about 62% for GPT-6 Astra, at a 200-fold lower token price by our arithmetic. Each of these results comes from the company making the claim.
As we read it, the business lever is where the money for specialising a model goes: today to forward-deployed engineers and closed-model tokens, which our AI token prices tracker lists per million; Trajectory is selling a platform for that work instead. Research on a model editing another model’s weights points at the same layer. That is a reading of the reports, not a forecast.
The Structural Read
As we read it, the funding is a bet on who does the specialising: TechCrunch describes the work Trajectory wants to automate as fine-tuning that often needs internal research teams or forward-deployed engineers.
As we read it, the sample is small: the August report itself says that at five tasks per category the resolution is coarse, so a gap of about two tasks is a thin margin on one benchmark.
As we read it, the cost claim matters as much as the pass rate: the June report says the post-trained Nemotron models matched the closed models’ quality at least ten times cheaper to run, plotted against price per token.
Ronak Malde, Trajectory co-founder, to TechCrunch
“But what the economy values is someone who worked on the job for 30 years”
Three Implications
FOR READERS WEIGHING THE ROUND TechCrunch reported $45 million at a $300 million valuation; The Information’s August figure, per Dealroom, was $40 million at $300 million post-money, and disclosed funding comes to $60 million by our arithmetic.
FOR READERS WEIGHING THE BENCHMARK The results are Trajectory’s own, on Harvey’s Legal Agent Bench, against Opus 4.6, Sonnet 4.6 and GPT 5.5; they have not been independently verified.
FOR READERS TRACKING TUNED OPEN MODELS Abacus.AI, Samaya and Palantir have each published a claim for a tuned Nemotron model since 24 October, each on a task the company chose to report.
The framework puts the role this way: “The emergence of forward-deployed engineering as a critical role in the AI industry reveals something fundamental about the current state of AI transformation: technical capability has far outpaced organizational readiness”.
As we read it, Trajectory is pitching software for part of that gap: TechCrunch says the fine-tuning it wants to automate often relies on forward-deployed engineers. The framework does not discuss Trajectory, and the reports do not show how much forward-deployed work a platform can replace.
The benchmark results are Trajectory’s own and have not been independently verified; the reports compare against Opus 4.6, Sonnet 4.6 and GPT 5.5, and we have not seen results for newer closed models on the same tasks. The 120-task count is our reading of the report’s five tasks per category across 24 areas. The two reports differ on Ultra’s base rate, 0% and 0.8%. Trajectory’s revenue, the terms of the round, and whether the August $40 million figure was an earlier close were not disclosed in the sources we read. We did not contact the company.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
Trajectory raised $45 million at a $300 million valuation, 2.6 times its reported seed valuation if both are post-money, by our arithmetic, to sell a platform that retrains open models on a customer’s own feedback. Its best published result, 8.3% against Opus 4.6’s 6.6% on held-out legal tasks, is about ten passes against eight out of 120, by our arithmetic.
A note on sourcing. Funding figures are from TechCrunch’s 2 November report and, for the earlier rounds, from Dealroom’s and Pulse 2.0’s summaries of The Information’s reporting and the seed announcement. Benchmark figures are from Trajectory’s own field reports and have not been independently verified. Task counts, multiples and totals are our arithmetic. We did not contact the company.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
