How Eventum cut LLM screening costs 104x with TypeSafe

TypeSafe now powers the structured decisions in our unattended recruiting pipeline. Across more than 154,000 job posts, it evaluates questions such as whether a role is US-only, whether it reflects a real hiring need, and how well an engineer fits the role.

This case study covers why we moved that layer away from frontier models, and the benchmark we ran to test whether the cost savings came at the expense of decision quality.

The challenge

The problem is less about model capability than volume.

A large share of our recruiting pipeline consists of repeated, bounded classification judgments. At frontier-model prices, those judgments become expensive at scale. The natural response is to screen fewer candidates, recompute less often, and reserve model calls for only part of the available data.

Generative models also introduce operational overhead. Structured outputs still need to be parsed and validated, creating a failure mode unrelated to the quality of the underlying judgment. And sampled outputs can move between runs, making numerical thresholds harder to treat as stable infrastructure.

The work

Typed outputs instead of parsed prose. TypeSafe supports yes/no judgments with calibrated probabilities, multiple-choice outputs, and rubric scores. Our code reads those outputs directly and branches on them: no JSON parsing, schema repair, or defensive parsing in the decision path.

One document, many questions. TypeSafe can evaluate dozens of prompts against a document in a single call. Our nightly enrichment process asks 44 questions of every job post in two API calls, so cost scales primarily with documents rather than criteria.

Reusable entity fingerprints. The same 30 typed questions describe every job, engineer, and sourced prospect. We compute those dimensions once, store them, and perform the first matching pass locally using cosine similarity. Only the strongest matches proceed to deeper paid evaluation.

Stable thresholds. In repeated testing, identical inputs returned identical probabilities to five decimal places. That allows us to use numerical thresholds for routing decisions rather than repeatedly invoking a generative judge.

Testing against frontier models

We evaluated the same 30 criteria using TypeSafe speed_latest, deepseek-v4-pro with reasoning disabled, and gpt-5.6-terra with thinking disabled.

The differences in cost and latency were substantial: TypeSafe was 30x cheaper than DeepSeek and 104x cheaper than GPT-5.6 Terra, while also being 30.5x and 21.9x faster, respectively.

We then ran a harder test on a platform-engineering role and 100 real engineer resumes. TypeSafe and DeepSeek produced direct fit scores for each candidate. Separately, GPT-5.6 performed hundreds of pairwise candidate comparisons, giving us a higher-fidelity ‘gold’ reference ranking.

TypeSafe’s ranking correlated strongly with the ‘gold’ reference ranking at Spearman ρ = 0.91, compared with ρ = 0.88 for DeepSeek.

Conclusion

Eventum has worked with TypeSafe as an early design partner and has run the model in production ahead of its public release. That gave us the opportunity to evaluate it against a real workload rather than a small set of demonstrations.

For the high-volume screening layer of our recruiting pipeline, the result was clear: TypeSafe delivered 30–104x lower cost and 22–31x lower latency, while preserving the ranking quality we needed.

The architecture that emerged is simple: fingerprint entities once with typed prompts, compare them locally where possible, and reserve more expensive model calls for the cases that survive.