TypeSafe AI has published a case study showing talent marketplace Jack & Jill replaced Gemini 3.1 Flash Lite with Jev for a high-volume candidate-ranking stage. The switch reportedly reduced model cost by 88 per cent and halved median processing time without a statistically significant loss in quality.

Jack & Jill uses AI agents to connect professionals and hiring teams. Its matching system evaluates about 6,000 candidates for each open role, produces an initial shortlist of 2,000 and then applies a more powerful model to identify a smaller group for human review.

The first stage created the cost pressure

Even with a low-cost Gemini model, applying an initial score to thousands of candidates for every role created substantial recurring expense. The company needed a replacement that could rank at scale while preserving the candidates hiring managers eventually wanted to meet.

This is the kind of narrow, repeated decision TypeSafe designed Jev to handle. The output is used by software as a structured ranking rather than presented as a conversational answer.

The comparison used 150 live roles

Jack & Jill tested Gemini and Jev on the same candidates across 150 roles. Each test set included every past winner—a candidate a hiring manager later requested—and a weighted sample of the remaining pool.

The key recall measure asked how many eventual winners each model preserved in the first 2,000-candidate shortlist. This is more relevant than a generic benchmark because it reflects the business cost of filtering out a suitable person before the detailed review stage.

Quality remained comparable

Jev retained 94.6 per cent of winners compared with 93.9 per cent for Gemini. TypeSafe notes that the 0.7 percentage-point difference was not statistically significant, with a 90 per cent confidence interval ranging from minus 0.4 to plus 1.8 points.

Jev also recorded an area-under-the-curve score of 0.933 compared with 0.924 for Gemini. That measure estimates how often the model ranks a winning candidate above someone with no recorded hiring-manager selection.

Median processing time was cut in half

The median time to screen candidates for a role fell from 20.3 seconds with Gemini to 10.3 seconds with Jev. Faster completion can make matching feel immediate and allows the product to run scoring in more places without creating a visible wait.

Latency should be monitored under production load as volumes and candidate profiles change. Median results are useful, but tail latency and failure rates determine whether every user receives the same experience.

Cost per thousand fell to nine cents

TypeSafe reports a cost of $0.092 per 1,000 candidates screened with Jev, down from $0.755 with Gemini. Its list-price comparison uses $0.042 per million input tokens with free output for Jev, against $0.25 input and $1.50 output for Gemini at the time of measurement.

At current volume, Jack & Jill projects annual savings of $265,000. Including expected growth, it estimates more than $500,000 over the next 12 months. Those are projections rather than audited realised savings and depend on volume and price assumptions.

The rollout took ten days

Within ten days of the first test, Jack & Jill moved the relevant stage to Jev for all calls. A drop-in change suggests the model fit the existing application contract without a long redesign.

Rapid migration should still include rollback, version pinning and ongoing comparison samples. Candidate distributions can drift, and a model that performs well on historical roles may behave differently as hiring demand changes.

Lower cost unlocked additional features

Jev now supports more than 15 workflows across the product. New candidates can be checked against suitable roles immediately after signup for under ten cents per person. A profile check reportedly fell from about 3.6 seconds to 0.2 seconds.

The company also uses Jev in its CV editor, company classification and marketplace filtering. Savings from those additional uses are excluded from the headline projection, according to the case study.

Hiring models require careful oversight

Screening affects access to employment, so cost and speed cannot be the only measures. Jack & Jill should continue testing performance across demographic groups, role types and changing labour conditions, with human review and a route to challenge incorrect results.

The case study provides unusually concrete production metrics for a specialised decision model. Jev’s strongest result is not simply being cheaper, but matching the measured baseline while enabling more frequent decisions. Long-term success will depend on preserving that quality and fairness as the system scales.

Buyers should interpret the results within the stated methodology. The quality test used historical hiring-manager selections as its signal, which can reproduce preferences or biases already present in the marketplace. Complementary audits should examine qualified candidates who were never selected and whether the ranking behaves consistently for new groups and roles.