Jev: The Cost-Effective AI for Decisions

The model is currently available for free on the Vercel AI gateway, allowing users to experience its capabilities without immediate cost.

The speaker highlighted its high utility across various domains, including personal productivity, coding, and product development, demonstrating its versatility and effectiveness.

Four cents per million input tokens represents the total cost for users, as the model generates minimal data and therefore charges no fees for output tokens.

This pricing model offers a significant cost advantage over standard LLMs, which typically impose substantial charges for generated output, thereby optimizing Jev for high-speed classification and decision-making tasks.

Jev excels in high-agency decision-making and logic gate applications, making it ideal for tasks such as routing feedback to support or classifying diverse data sets.

In contrast, standard LLMs are better suited for generative tasks, including drafting text or writing code, while Jev's speed allows for efficient real-time application processing.

Jev delivers three primary output types: a choice, a score, or a null, fundamentally differing from text-based generative models.

A 'choice' selects from a predefined list, a 'score' provides numerical rankings (e.g., severity levels), and a 'null' (Boolean) response indicates the probability of a 'yes' answer to a specific question.

Jev is used to classify unstructured data, a process previously challenging and costly to organize.

The speaker leveraged Jev to perform pairwise comparisons on thousands of PRs, assessing their thematic relatedness.

This method enabled granular tracking of engineering effort across various initiatives and cost only 9 cents for 1,700 PRs (17,000 comparisons), completed in approximately 2 minutes.

Leadership can answer board-level questions regarding effort allocation by using this tool, which performed a classification of 112 pull requests for 1.1 cents and analyzed a project of 2,000 pull requests for 9 cents, with the speaker noting that nearly 30% of work at ChatPRD is dedicated to platform, security, and infrastructure.

This approach helps leadership address board-level inquiries regarding effort allocation by providing clear data.

The analysis revealed that nearly 30% of the work at ChatPRD is dedicated to platform, security, and infrastructure, offering critical insights into resource distribution.

Jev provides fast and accurate categorization of investment areas such as security, performance, and infrastructure, offering a clear picture of resource allocation.

The speaker observed shifts in investment priorities over several months, demonstrating the model's utility in tracking evolving strategies on large datasets at minimal cost.

Parsing desktop data from local Claude and Codex chat logs enables users to track granular changes in their daily effort, such as the shift from pure engineering tasks in January to agent-based work and media publishing by September.

Data showed a significant shift from pure engineering tasks in January to more agent-based work and media publishing by September, providing granular insights into daily effort distribution at almost no cost.

Jev, when paired with an LLM, proves highly effective for automating email triage, categorizing and scoring emails for deletion in personal Gmail accounts.

Jev functions as a filter for tagging and clustering information, while the 'LLM buddy' handles higher-level reasoning, creating a powerful pattern for managing high-volume data like bugs and emails.

The speaker successfully leveraged Jev for classification, clustering, and mapping, strategically reserving advanced models like Astra for strategy and insight extraction.

Success is attributed to applying Jev specifically where structural processing is needed, rather than relying solely on large generative models for all tasks.

Jev assigns labels to various data signals, including PRs, support tickets, and linear tickets, effectively synthesizing over 1,100 individual signals to identify discrepancies between customer requests and actual development work.

Astra is then paired with Jev clusters to extract deeper insights from the labeled data, enhancing the analysis.

This approach has made the feature more margin-accretive by replacing labor-intensive manual analysis, with the pipeline performing over 200,000 classifications and groupings for approximately four dollars on the Jev side.

A live audience dashboard was developed to analyze 4,500 YouTube comments from the How I AI podcast, with Jev classifying sentiment (positive, negative, neutral) and identifying specific episode requests, providing clear metrics on audience feedback.

The architecture employs batching, scoring, and prioritizing high-score results to maintain performance, allowing for quick filtering of specific topics like 'screen share' or 'slop'.

An application that maps voice sentiment to visuals and quotes was built, integrating OpenAI's Realtime Voice API, Jev, and an external Quote API.

Jev classifies the user's emotion and maps it to a hex color code, filtering quotes based on the identified sentiment and displaying them in real-time with manageable latency and potential for further optimization through caching.

Jev analyzes the speaker's emotional state from voice input to determine a corresponding color, highlighting the potential for building interactive, real-time app experiences.

The system employs a predefined list of hex values for color matching, with Jev scoring colors based on incoming phrases from the real-time API.

The model filters quotes against detected sentiment before displaying them, and local execution can be further optimized through caching to enhance performance.

The host announced plans for another Jev-focused episode featuring a special guest and encouraged viewers to submit questions regarding Jev implementation in the comments section.