Benchmarking Coding Agents on New vs Legacy Codebases — Denys Linkov, Wisedocs
AI Engineer World's Fair 2026 · 18:08
AI claims document analysis
Wisedocs develops AI software for complex casualty claims, helping insurers, government agencies, defense counsel, and medical experts organize, analyze, and summarize medical, legal, and billing records. Its platform combines domain-specific models with expert human oversight to turn document collections into structured information for claims decisions. The Billing & Financial Analyzer identifies cost drivers, billing anomalies, and billed, paid, and outstanding amounts. The Deposition Analyzer connects testimony to medical and factual records, flagging contradictions for legal and claims teams.
Co-founders Connor Atchison and Jenna Earnshaw built Wisedocs around problems Atchison encountered in military healthcare administration and family caregiving. Earnshaw became CEO in 2026, while former CEO Atchison became President, responsible for product, operations, and infrastructure. Its research includes MLCR, a benchmark and evaluation harness that tests reasoning across long medical records. By inserting unrelated documents among relevant evidence, it measures how well models maintain accuracy when claim files contain distracting material.
In June 2026, the company reported processing more than 130 million claims documents and an 84-fold increase in enterprise revenue over the preceding 24 months. Its financing includes C$4.5 million in growth capital provided by CIBC Innovation Banking in 2024 to broaden its client base and expand its product offering.
AI Engineer World's Fair 2026 · 18:08
Affiliations reflect their AIE appearances, not necessarily current employment.
Learn how Linkov compared agent-assisted refactoring across model generations in a legacy, multi-repository pipeline.
Denys LinkovAI Engineer World's Fair 2026
Start here for practical model-training requirements and the inner and outer team loops that shape cross-functional AI work.
Denys LinkovAI Engineer World's Fair 2025
Linkov discusses orchestration evaluation and Temporal prototyping alongside coding-agent reliability and task-duration limits. Reported outcomes include lower processing costs and times, larger-file support, and faster feature delivery.
Linkov argues that adaptable generalists and existing teams with domain experience can outperform reflexive hiring of specialist AI researchers. He places continuous learning and delivery bottlenecks at the center of team design.
Affiliations reflect each recorded session, not necessarily current employment.