# Mission-Critical Evals at Scale Christopher Lovejoy, Anterior | AI Engineer Summit 2025 | 12:15 Source: https://www.youtube.com/watch?v=cZ5ZJy19KMo Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks. Page: https://aietalks.com/talks/mission-critical-evals-at-scale Published: 2025-02-22 Tags: evals, healthcare, human-in-the-loop ## TL;DR - Human review alone cannot keep pace with a medical AI system processing tens of thousands of decisions each day. - Reference-free evaluations estimate correctness before human labels exist, so teams can monitor every case and prioritize the cases most likely to be wrong. - A feedback loop between automated evaluation and targeted human review can improve accuracy while keeping the clinical review team small. ## Summary Christopher Lovejoy explains how Anterior evaluates AI-generated medical prior authorization decisions as its product scales. Human review produces useful critiques and ground truths, but reviewing a fixed percentage of cases quickly becomes too expensive. Offline datasets also miss new edge cases until after they have reached customers. Anterior addresses both problems with real-time, reference-free evaluations. An LLM judge, confidence estimation, and logic-based checks estimate whether each output is correct before a human review exists. These scores help measure performance across all cases, compare automated evaluations with human judgments, and prioritize cases for clinical review. Over time, human reviews validate the evaluator and expose new failure modes. Lovejoy says Anterior reviews tens of thousands of cases with fewer than 10 clinical experts and reported an F1 score of nearly 96% in a recent prior authorization study. The system can also route low-confidence cases to a more expensive model, an internal clinician, or the customer's review queue. ## Key ideas ### Scale creates medical edge cases that an MVP can hide [00:31](https://www.youtube.com/watch?v=cZ5ZJy19KMo&t=31s) Lovejoy says an LLM product can reach an MVP quickly, but higher request volume exposes inputs and failure cases that the team has never seen. At Anterior, the system helps decide whether a treatment request should be approved or reviewed by a clinician. One example produced an answer that called evidence of multiple sclerosis "suspicious" even though the patient already had a confirmed diagnosis. That distinction changes the medical meaning and made the answer wrong. A mistake that appears once in 1,000 or 10,000 cases becomes a large operational problem when the system processes more than 100,000 cases each day. ### Human review produces valuable labels but does not scale by percentage [02:13](https://www.youtube.com/watch?v=cZ5ZJy19KMo&t=133s) Anterior's clinical reviewers inspect the medical record, guideline, question, and required context in an internal dashboard called Scalp. They can label an answer as incorrect and write a critique explaining what went wrong. Those critiques can then be combined with the original answer to generate a description of the correct answer, or a ground truth, for offline evaluation. Lovejoy gives a simple scaling example: reviewing half of 1,000 daily decisions requires 500 reviews, or five clinicians at 100 reviews each. At 10,000 decisions, the same percentage requires 50 clinicians. Reducing review to 5% only postpones the problem when volume reaches 100,000 decisions. ### Offline evaluation sets are useful but arrive after new failures [04:34](https://www.youtube.com/watch?v=cZ5ZJy19KMo&t=274s) Offline datasets can contain gold-standard cases generated from human reviews. Teams can split them by enterprise, medical condition, difficult question, complex case, or ambiguous outcome, then track performance over time while changing the AI pipeline. Lovejoy says the danger is timing. If a new edge case is added only after it has appeared in customer data and received a human review, the system may already have made the mistake in production. Medical records have a highly heterogeneous input space, so new cases continue to appear at scale. Offline evaluations alone cannot provide an early warning for every failure. ### Reference-free evaluation estimates correctness before a human label exists [05:32](https://www.youtube.com/watch?v=cZ5ZJy19KMo&t=332s) A reference-free, or label-free, evaluation runs before the true outcome is known. An output from the AI pipeline is passed to an evaluator with a scoring system. The evaluator might judge helpfulness, concision, tone, or confidence in a binary or multiclass result. Anterior's output is either approval or escalation for review. Its reference-free system produces a confidence grade ranging from high confidence that the answer is correct to confidence so low that the answer is considered wrong. A threshold can turn that grade into a predicted correct output, allowing the system to estimate performance while decisions are still being processed. ### Automated scores can cover every case and guide human sampling [07:19](https://www.youtube.com/watch?v=cZ5ZJy19KMo&t=439s) Reference-free evaluations give Anterior an estimated performance measure across all incoming medical decisions, including cases that no clinician reviews. For cases that do receive human review, the team compares the automated result with the human result to measure alignment and decide how much to trust the evaluator. Confidence grades can also be combined with context such as procedure cost, bias risk, and previous error rates. The system then prioritizes cases with a high probability of error or high consequence for human review. This makes the review set change with the system's observed behavior instead of staying at a fixed percentage. ### Human review validates the evaluator and improves its ability to find new errors [08:27](https://www.youtube.com/watch?v=cZ5ZJy19KMo&t=507s) Lovejoy describes the process as "validating the validator." Reference-free evaluations surface cases, and human reviewers determine whether the suspected problem is real. Those reviews improve the system and help it handle future edge cases. Over time, the number of unfamiliar edge cases should fall while detection improves. Lovejoy also argues that this process is difficult for a competitor to copy because it depends on processing high volumes of real data and running repeated data-driven iterations. The advantage comes from the accumulated evaluation system, not only from the original product pipeline. ### Low-confidence outputs can trigger different levels of intervention [08:57](https://www.youtube.com/watch?v=cZ5ZJy19KMo&t=537s) Once the evaluation system is trusted, it can sit inside the production pipeline. A confident output can go directly to the customer. A low-confidence output can be sent through another LLM pipeline with more expensive models, routed to an internal on-call clinician, or placed in the customer's review dashboard. This lets the team match the amount of checking to the estimated risk of each case. The system can therefore protect customer-facing decisions without sending every decision to a human or applying the most expensive processing to every request. ### Anterior says targeted review supports a small clinical team [09:40](https://www.youtube.com/watch?v=cZ5ZJy19KMo&t=580s) Lovejoy says Anterior has avoided hiring an ever-growing group of clinical reviewers. He contrasts its team of fewer than 10 clinical experts, which reviews tens of thousands of cases, with a competitor he says has hired more than 800 nurses for reviews. After several iterations, Anterior reported strong alignment between its AI and human reviews, at a level comparable to the alignment among its human reviewers. He also says the system helped the company respond quickly to errors and meet customer time expectations. A recent study reported a prior authorization F1 score of nearly 96%. ### The system should be built around live data and high-quality review [10:51](https://www.youtube.com/watch?v=cZ5ZJy19KMo&t=651s) Lovejoy recommends treating review data as input for building, auditing, and improving the evaluation system itself. He says teams should evaluate live production data rather than depend only on offline datasets, so problems can be identified while they are occurring. He also recommends choosing the best reviewers, prioritizing review quality over review volume, and building internal tooling when it helps the team move faster. At Anterior, these choices produced real-time performance estimates and a review process designed to scale with demand. ## Notable quotes - "A mistake that might happen every thousand cases or even every 10,000 cases, but if you're processing more than 100,000 cases every day then that's a lot of mistakes that you need to pick up." (01:41) - "Reference free also known as label free means that you evaluate before you know the true outcome." (05:32) - "Our reference free evals surface the cases and then our human review determines the accuracy." (08:27) - "Don't just use review data to audit your performance, use it to build, audit and improve your auditing system, your evaluation system." (11:11) ## Tools & references mentioned - Anterior - Scalp - Florence - LLM as judge ## Who should watch - You run an LLM product in healthcare or another setting where a rare wrong answer can create serious consequences. - Your current evaluation set is built from reviewed cases, but production volume keeps producing unfamiliar inputs. - You need to reduce manual review while still deciding which outputs deserve a clinician's attention. ## Related talks - [Engineering Better Evals: Scalable LLM Evaluation Pipelines That Work](https://aietalks.com/talks/engineering-better-evals-scalable-llm-evaluation-pipelines-that-work) (Dat Ngo & Aman Khan, Arize, 24:46) - [Evals-Driven Development for a Mental Health AI Coach](https://aietalks.com/talks/evals-driven-development-for-a-mental-health-ai-coach) (Akele Reed, Dave Revere & Doug Keller, SonderMind, 21:17) - [Mastering AI Evaluation: From Playground to Production](https://aietalks.com/talks/mastering-ai-evaluation-from-playground-to-production) (Doug & Carlos Essan, Braintrust, 1:25:08) - [Build Evals That Actually Matter](https://aietalks.com/talks/build-evals-that-actually-matter) (Nick Ung & Akshay Sharma, Lyft, 37:45) - [How to Run Evals at Scale: Thinking Beyond Accuracy or Similarity](https://aietalks.com/talks/how-to-run-evals-at-scale-thinking-beyond-accuracy-or-similarity) (Muktesh Mishra, Adobe, 09:25)