Jacob Lauritzen is the chief technology officer of Legora, where he leads engineering and technical strategy for a collaborative legal-AI platform. Previously the founder of companies serving investors and sales teams, he now builds systems that give lawyers meaningful control over automated research, contract review, and drafting.
From company intelligence to legal AI
Lauritzen founded Firmnav to help private-equity investors and mergers-and-acquisitions advisers identify investment targets and comparable businesses using data mining, natural-language processing, and machine learning. After selling the company, he adapted that experience to business-to-business sales, founding Unhaze, a Y Combinator Summer 2023 company. Unhaze combined historical deal information with signals including hiring, organizational changes, technology choices, and company activity to identify promising prospects. Lauritzen described the transition between the two ventures as applying company-intelligence technology to a new commercial problem.
He subsequently joined Legora as head of engineering and advanced to chief technology officer. During that period, the engineering organization grew from approximately 10 people to 40. He favors hiring engineers with founder-like ownership and agency, extending the entrepreneurial orientation of his earlier ventures into a growing technical organization.
Verification determines the limits of automation. Checking contractual definitions is comparatively straightforward; judging whether an agreement reflects sound commercial strategy, or whether litigation will succeed, is not. Lauritzen assigns checkable operations to agents while reserving risk posture, negotiation strategy, and precedent selection for professionals. Trusted documents provide useful proxies when outcomes cannot be conclusively verified in advance.
Human judgment belongs inside the workflow. He structures complex assignments as connected subtasks, with reusable skills capturing domain-specific requirements and contingencies. Bounded permissions establish trust, while an agent decision log lets systems record provisional choices for subsequent review without stopping whenever uncertainty arises.
Documents are better collaboration surfaces than endless chat. Lauritzen advocates moving beyond linear conversational interfaces toward durable collaborative artifacts: editable contracts and tabular reviews where lawyers can inspect individual clauses, flag exceptions, delegate specific tasks, and intervene directly.
Evaluate the deployed system, not only the model.Legora’s Benchmark for Agentic Reasoning assesses legal work across the complete combination of models, tools, documents, skills, and application environment. Its broader corpus contains 5,161 cases across 28 practice areas; the benchmark draws on a representative subset and assesses factual grounding, analysis, citations, and recommendations.
When agents can produce complex work cheaply, the hard part becomes steering and reviewing it. Persistent documents and tables put human judgment where the work happens.
Legora’s migration from Elasticsearch through Postgres to turbopuffer shows how project lifecycles, customer encryption requirements and storage round trips shape a search engine for billions of legal documents.
Roughly 4,000 partitions still mixed active and inactive projects. One namespace per project let cold collections stay in object storage without sharing a large partition with active work.
Direct object-storage writes accept hundreds of milliseconds of latency. Responsive reads require indexes and query plans that do substantial work within a few storage round trips.
Customer encryption requirements can include SSD caches. Disabling the disk cache performed well enough for some Legora workloads, avoiding the need to implement encrypted caching first.
Legal research fans out because jurisdictional hierarchy, temporal validity and regulatory exceptions require additional retrieval. The application must resolve those relationships beyond finding similar passages.
Cluster trees let frequently consulted vector-routing information stay in DRAM while larger leaf data occupies cheaper tiers. Full-text search also needs to control posting-list transfers and memory-bandwidth costs.
The economic fit depends on a long tail of cold collections and tolerance for occasional cold-read delays. For Legora, that fit reduced infrastructure work and freed engineering time for the product.
