Kev — Train and run small, specialized AI decision models on your own hardware for classification.
Analyzed by Sai Pavan Gopularam · AI · Decision Models · View on GitHub
- Stars: 5538
- Forks: 296
- Commits last 30 days: 100
- Health: Active (100 commits this month)
- Language: Python
- License: Apache-2.0
What It Is
Kev is a family of small, specialized AI models designed to make specific decisions from text, similar to how a human expert would categorize information. Think of it as a highly focused AI assistant that can answer yes/no questions, pick from multiple choices, or rate something on a scale, all based on a piece of text you provide.
This matters because traditional large language models are often too general, expensive, or slow for repetitive, high-volume classification tasks. Kev solves this by offering compact models that run locally, providing fast, consistent, and auditable decisions for things like customer support ticket routing or content moderation.
License Verdict
Apache-2.0 License — Build and Sell Freely — Commercial Use Approved • Permissive Open Source
The Apache-2.0 license is highly permissive. You can use, modify, and distribute this software for private or commercial purposes. You must include a copy of the license and retain all copyright and patent notices. There are no copyleft restrictions, meaning you don't have to open-source your modifications.
How to Use It
Set up Kev locally to run decision models via an API. It involves cloning the repository, installing dependencies with `uv`, and starting the server. You can then interact with it via `curl` or the Python SDK.
Prerequisites:
- Python 3.12+
- uv
- Node 20.9+ (for playground)
Estimated setup time: 10 minutes.
git clone https://github.com/jaredpalmer/kev.git && cd kev
uv sync --extra serve
uv run --extra serve python -m kev.serve --run jaredpalmer/kev-4b --port 8009
curl -s localhost:8009/v1/systemone -H 'content-type: application/json' -d '{ "state": "Example text", "model": "kev-latest", "questions": { "is_urgent": {"type": "noul", "instructions": "Is this urgent?"} }}'What I'd Build With This
Email Triage for Small Businesses (micro-saas)
A simple web app where small business owners connect their customer support email inbox. Kev automatically categorizes incoming emails (e.g., "Refund Request," "Technical Issue," "General Inquiry") and assigns an urgency score. This streamlines initial handling, letting owners focus on critical issues.
Effort: 1 Week Build Time · Target: Small Business Owners · Pricing: $29/month
AI-Powered Policy Compliance Checker (saas)
A platform for mid-sized companies to upload internal documents (contracts, HR policies, user agreements). Kev analyzes these documents to identify clauses that violate specific compliance rules or highlight potential risks, flagging them for human review. This automates a tedious and error-prone process for legal or HR teams.
Effort: 3 Months Build Time · Target: Mid-Market Legal/Compliance · Pricing: $299/month per team
Internal AI Agent for Customer Service Escalation (enterprise)
An on-premise or private cloud deployment that integrates with an enterprise's existing CRM/helpdesk system. Kev acts as a first-line AI agent, analyzing customer interactions (chat logs, call transcripts) in real-time to determine if an issue requires immediate human intervention, which department to route it to, and the customer's sentiment. This reduces human workload and improves response times.
Effort: 6 Months Build Time · Target: Large Enterprise CX Teams · Pricing: $5,000+/month (usage-based)
Sai Pavan Gopularam's Take
Kev is a really interesting project for anyone needing fast, local, and specialized text classification. Instead of fine-tuning a massive LLM for every niche task, you get a small, efficient model that can run on consumer hardware. I'd lean into building a highly specialized customer support bot for a niche industry, charging $99/month for automated ticket routing and sentiment analysis.
Watch Out For
- Option Order Matters: While questions are isolated, the order of options within a single 'choice' question can still influence the model's answer. Test permutations if this is critical for your use case.
- Apple Silicon Performance: On Apple Silicon, Qwen3.5 models run via MLX, not PyTorch, due to a lack of PyTorch kernels for specific layers. This means performance characteristics will differ from CUDA/ROCm.
- Exact Evaluation Path: For precise replication of evaluation results, ensure you set `KEV_DTYPE=fp32` when serving, as the default `bf16` is a latency optimization that might slightly alter probabilities.
- Date Calculations: Kev doesn't inherently subtract dates well. Enable `KEV_DATE_FACTS=1` to append date differences to the state for improved accuracy on date-sensitive policy questions.
I break down trending repos like Kev every week — join the newsletter.