Terminal X uses its Capital Memory system to encode how a firm's senior investors make decisions into deployable agents, and it evaluates those agents’ outputs with an Eval Agent that runs on GPT‑6.1 Sol and other OpenAI models. The Eval Agent scores every output against a 54‑criterion rubric to check for mandate compliance and other firm standards.
How the solution is built
- Terminal X’s engineers — many former JPMorgan and Blackstone finance professionals — spend weeks onsite with each customer to reverse‑engineer written and unwritten investment processes (for example, how a credit desk screens borrowers or which sources a co‑investment team consults before drafting a memo).
- Those practices are encoded into hundreds of customized agents by asset class and strategy, with sub‑agents for each stage of the investment workflow.
- GPT‑6.1 Sol handles judgment‑intensive steps that need a very large context window (synthesizing the investment thesis, identifying key considerations, and testing fit with the fund mandate). GPT‑6 Luna runs higher‑volume steps where token efficiency is more important than depth.
- Agents execute inside each firm’s private data room, process documents in English, Korean, Japanese, and Chinese, and store what they learn as a versioned firm Memory: a record of how a desk makes a particular kind of decision, including past analyses, committee votes, workflow instructions, and house views.
- Every production query is auditable: Terminal X records the steps taken, sources retrieved, model used, and the reasoning trail.
Measured results with GPT‑6.1 Sol
- The Eval Agent evaluates outputs against a 54‑point rubric. Using GPT‑6.1 Sol, it detected roughly 25% more recommendations that departed from a fund’s mandate versus the next‑best model.
- Agreement between the Eval Agent and senior human reviewers was 88.6% on GPT‑6.1 Sol, versus 86.4% for the next‑best model — a 220 basis‑point improvement.
- GPT‑6.1 Sol’s token efficiency on long‑context tasks enabled a deeper review workflow without increasing budget: the thesis is first tested against the fund mandate, then compared to recent deals; uncertain cases are rerun and reconciled before reporting.
- That deeper review caught deals that a single pass on a comparable model had allowed through. GPT‑6.1 Sol also ran the same workflows at about 30% lower cost per query than the next‑best model, at similar speed.
Usage and savings
- One private equity customer reported $2.4 million in annualized savings from automating due diligence and investment memo workflows.
- To date, Terminal X has built 1,700 firm Memories, executed more than 1 million custom workflows, and reached weekly active user figures equal to 81% of its monthly active users.
How evaluation feeds back into the system
- The Eval Agent’s 54 checks ask whether numbers, comps, consensus figures, and quotes come from the right sources; whether the narrative holds together; and whether a recommendation ties back to the house view.
- Each check is graded pass or fail. Failures come with a one‑line reason and a suggested corrected text or better source; those failures are then encoded as new workflow instructions or test cases in the relevant Memory (for example, a missed source, a wrong comparable, or a thesis that drifted from house view).
- As Ken Lee, Terminal X’s co‑founder and CTO, puts it: “Every answer gets eval’d, and whatever the analyst corrects teaches the system a little more about how that firm works. So the longer a firm uses it, the more it thinks like their own team.”
Next steps and vision
Terminal X plans to extend firm Memories from desks to individual investment professionals so that each person’s actual practiced standard—not just the described standard—is captured. Hyun, Terminal X’s co‑founder and CEO, frames the goal simply: “The firms that win the next decade will own their judgment as data. Our job is to make that judgment measurable, so it can be kept, improved, and reused.”
