Personal AI

Train an AI on how you work.

Agent skills for turning selected AI chats and Screenpipe context into a personal or company model with River.

Your conversations contain preferences. Your work history contains examples. This toolkit helps an AI coding agent turn the parts you choose into reviewed training data, run a small fine-tuning experiment, and test what changed.

Explore Screenpipe Enterprise to scope a team pilot, data access, and implementation support. Training, inference, and managed delivery are scoped separately. This repository is a self-directed experimental toolkit.

Requires Python 3.12 or newer for River execution. Local preparation and dry-runs use only the standard library.

Start with your agent

Clone this repository, open it in Codex, Claude Code, or another coding agent, and paste:

Read skills/curate-work-context/SKILL.md. Help me prepare a small dataset from AI chats and Screenpipe history I select. Keep it local while we review it. Then use skills/train-with-river/SKILL.md to plan a River experiment and skills/evaluate-personal-model/SKILL.md to compare the results. Start with one behavior I want my model to learn.

git clone https://github.com/screenpipe/river-ai-screenpipe-training.git
cd river-ai-screenpipe-training
python3 scripts/prepare_dataset.py examples/reviewed.jsonl --out data/demo
python3 scripts/train.py --data data/demo

The last command validates the data and prints a local plan. It makes no API calls and does not download a model. The included examples are synthetic plumbing fixtures, not enough data to produce a useful personalized model.

Keep the skills in this checkout so their linked documentation and scripts remain available. Reference their paths directly in your agent; no global installation is required.

The workflow

Skill	What it produces
Curate work context	Reviewed prompt/answer examples with provenance and an independent evaluation split
Train with River	A bounded LoRA run, baseline outputs, a saved checkpoint, and sampled outputs
Evaluate a personal model	A comparison against the same prompts, failure cases, and a serving decision

Sources can include selected Codex or Claude conversations, user-provided ChatGPT exports, reviewed Screenpipe OCR/transcripts, and written procedures. The skills use your agent's existing local tools. This repo does not silently crawl your computer or upload raw history.

Run a real experiment

1. Follow the curation skill and replace the synthetic examples with reviewed examples. See the data format.
2. Install dependencies in a virtual environment: python3 -m venv .venv, activate it, then pip install -r requirements.txt.
3. Configure RIVER_API_KEY through your environment or secret manager. Never put it in a command argument, dataset, or committed file.
4. Check River model access and choose a model available to your account.
5. Review the local plan, then run with explicit execution:

python scripts/train.py --data data/your-project --base-model YOUR_AVAILABLE_MODEL \
  --steps 10 --max-seconds 300 --execute

The execution sends the reviewed prompt/completion tokens and evaluation prompts to River and can incur training/inference charges. The time limit is a client-side stop between requests, not a provider billing cap; an in-flight request can outlast it. Runs do not retry automatically. Outputs stay in ignored runs/.

What this does and does not show

Fine-tuning can teach recurring response patterns, language preferences, and procedures. It is not a reliable database for changing facts. Use retrieval for source-grounded recall and evaluate both approaches for your task.

A small demo is not a personal superintelligence, a proven company clone, or evidence of broad reliability. Keep unrelated conversations out of training, hold out entire source groups, and test failure cases. Passing a few training-like questions does not establish generalization.

A saved River checkpoint is separate from a deployed endpoint. Confirm model and serving support in your River account before promising a live chat UI. This toolkit does not enable paid serving or change account configuration automatically.

Validation

python3 -m unittest discover -s tests -v

See validation coverage. Local tests exercise preparation, leakage rejection, masking and dry-run behavior. A successful local test is not a completed River training run.

Built by Screenpipe, using River's public API. No official River partnership is implied.