sploink 0.2.0

Auto-discover repetitive tasks in your AI agent and transparently replace them with small, specialized, locally-trained models.

sploink

Auto-discover the repetitive tasks inside your AI agent, and transparently replace the expensive ones with small, specialized models trained on your own real traffic.

Status: Early. The connect/discover/route mechanism works locally, end to end, against Ollama and a local mlx_lm.server on Apple Silicon. No hosted version yet, no Linux/cloud GPU training path yet.

Why

Most of what an AI agent does day to day is the same narrow, repetitive task, over and over -- classify this, check that, draft a reply -- but every call goes to one expensive, general-purpose model regardless of whether the task actually needs that much capability. sploink finds those narrow tasks automatically, from your agent's real traffic, and lets you train a small local model specifically for one of them -- then transparently routes future calls to it, with no changes to your agent's code beyond one line.

Install

# Core install -- just the connect/discover/route layer
pip install sploink

# With local LoRA training support (Apple Silicon only, via mlx_lm):
pip install "sploink[train]"

Quickstart

The only integration line you ever write, regardless of what else you do:

import sploink
sploink.init(account_id="your_account")

# ...the rest of your existing agent code, completely unchanged...

This patches the OpenAI-compatible client your agent already uses. From that point on:

• Every call gets logged.
• The real tools=[...] schema on each call is read to discover the distinct tasks your agent actually performs -- no source-code parsing, works with any framework built on the openai Python SDK against an OpenAI-compatible backend.
• If a task already has a trained adapter, matching calls are transparently routed to it instead of the generalist -- including bundled, multi-tool ReAct-style calls, split into a cheap "which tool" decision and a routed "what arguments" call under the hood.

Training a task (requires the train extra):

adapter_path = sploink.create_lora_adapter(
    task_name="classify_expense",
    system_prompt="...",
    train_examples=[(text, {"category": ..., "amount": ...}), ...],
    valid_examples=[...],
)

Once trained, the very next real call for that task routes to it automatically -- no further code change.

What this isn't (yet)

• No hosted/cloud version -- everything here runs on your own machine.
• No Linux or non-Apple-Silicon training path (mlx_lm is Mac-only).
• No automatic quality gating before a newly trained adapter starts serving real traffic -- that's on the roadmap, not yet built.