SAI - Sri's Personal AI

A Self-Improving AI Assistant That Reads Research and Evolves. Built on OpenClaw, running 24/7 on a Hetzner VPS. SAI ingests AI research papers, extracts actionable insights, and proposes changes to its own operating rules - with human approval at every step.

What SAI (Sri's Personal AI) is and what it does.

SAI (Sri's Personal AI) is a personal AI assistant that doesn't just respond to requests - it actively learns from AI research and proposes improvements to how it operates.

Every change goes through explicit human approval. It's an experiment in building an AI system that gets better over time without losing human oversight. This is what happens when you take a 10-day AI course and keep building.

Beyond self-improvement - SAI is a working personal assistant used daily.

Self-improvement is the most technically interesting part, but SAI is also a daily-use assistant. Here's what it does on a regular basis:

CapabilityHow It Works
AI News DigestsDaily Telegram messages summarizing the latest AI developments
Research Paper SummariesCron job extracts key findings from ingested papers and sends condensed briefings
Google CalendarReads upcoming events and provides schedule-aware context in conversations
Email AwarenessReads and triages Gmail, sends notifications for important messages
Web ResearchSearches the web and reads pages on demand via browser automation
Self-ImprovementReads AI research, proposes operating rule changes, waits for human approval

The self-improvement loop gets its own deep-dive below - but it's one part of a broader system that's useful every day, not just when it's learning.

System architecture diagram and data flow explanation.

Two cron-driven pipelines work together: one ingests research at zero LLM cost, the other reasons over the knowledge base and proposes changes.

📥
Research Fetcher Linux cron · every 6h · free
Round-robin across sources with smart query generation from AGENTS.md
arXiv RSS HuggingFace Semantic Scholar
🧠
Learning Report OpenClaw cron · every 3 days · ChatGPT Plus
Reads ChromaDB, reasons over ingested papers, proposes concrete changes to agent config
ingest
propose
💾
ChromaDB Vector store · on-disk · local embeddings
Growing knowledge base · 4 tiers
Hybrid search: semantic + keyword (BM25)
Pinned Hot Warm Cold → auto-prune
📲
Telegram Approval Human-in-the-loop · per-diff
“approve all”, “approve 1,3”, or “skip”
Each diff shown individually
Backup + security audit after every write

The research-fetcher runs every 6 hours as a native Linux cron job, round-robin across three sources with automatic deduplication. The learning-report runs every 3 days, reads the vector store, reasons over ingested papers, and sends proposed changes to Telegram for per-diff human approval.

SAI reads research, proposes config changes, and waits for human approval on every diff.

SAI reads research papers, extracts techniques, and maps them to concrete changes in its own configuration. Every proposed change requires explicit approval - SAI shows the exact diff, waits for a "yes", and runs a security audit after applying.

Example: Learning from Self-RAG

After ingesting the Self-RAG paper (Asai et al., 2023), SAI proposed adding a retrieval decision gate: "Before any knowledge-base lookup, decide explicitly whether retrieval is actually needed." This was reviewed, approved, and applied as a 4-line change to AGENTS.md.

Result SAI now skips unnecessary searches for drafting and creative tasks, reducing token usage and improving response relevance.

6 Approved Self-Improvements

ChangeDerived From
Selective retrieval gateSelf-RAG
Self-critique before sendingSelf-RAG + Chain-of-Note
Separate evidence from inferenceRAG Survey
Coverage check on retrieval answersRAG Survey
Templated approval protocolRAG + Self-RAG
Validated-only memory updatesSelf-RAG

5-layer progressive trust model with human approval at every critical step.

The system uses a 5-layer progressive trust model. The first two layers can't change SAI's behavior - they only add data and generate reports. From layer 3 onward, nothing happens without a human in the loop.

LayerWhat It DoesChanges Behavior?Human Gate
1. IngestStores papers in vector DBNo--
2. SynthesizeGenerates a learning reportNo--
3. ProposeSends proposed changes via TelegramNoYou read it
4. ApplyModifies config filesYesPer-diff approval
5. Skill CreationDrafts a new skillYesManual review + scan
Zero unsupervised behavior changes Layers 1–3 run automatically but can't modify how SAI operates. Layers 4–5 can - and both require explicit human approval before anything is applied.

Multi-source fetcher, smart queries, and knowledge base maintenance tiers.

Multi-Source Research Fetcher

Round-robin across three sources with automatic deduplication. Runs as a native Linux cron job - zero LLM cost.

arXiv RSS HuggingFace Daily Papers Semantic Scholar API

Smart Query Generation

Instead of generic searches, the fetcher reads SAI's current AGENTS.md and MEMORY.md to derive targeted research queries. If SAI recently added a rule about "self-critique", the fetcher searches for papers on LLM self-critique methods.

Knowledge Base Maintenance

Tier-based lifecycle management keeps the vector database clean as it grows.

Pinned - led to approved changes, never deleted Hot - last 30 days Warm - 30–90 days Cold - 90+ days, auto-pruned

Technology stack and hosting setup.

SAI runs on a ChatGPT Plus subscription for the AI engine and a shared Hetzner VPS for hosting. Everything else - vector storage, embeddings, parsing, fetching - is free and self-hosted.

ComponentTechnology
AI EngineOpenClaw on Docker (ChatGPT Plus)
HostingHetzner VPS (shared with other services)
Vector DBChromaDB (on-disk)
Embeddingsall-MiniLM-L6-v2 (local)
PDF ParsingDocling
Paper FetchingLinux cron + Python
ChannelTelegram Bot
SearchBrave API (free tier)
MonitoringUptime Kuma + Glances

Beyond the Course - everything built after completing the 10-day OpenClaw Mastery course.

This project started as the OpenClaw Mastery For Everyone 10-day course. Everything below was built after completing the course.

Day 11

Teach It to Learn

RAG pipeline with ChromaDB, Docling PDF parsing, sentence-transformer embeddings, hybrid search skill, and human-gated self-improvement loop.

Day 12

Multi-Source Research Fetcher

Round-robin fetcher across arXiv RSS, HuggingFace, and Semantic Scholar. Smart query generation from agent config. Zero-cost Linux cron.

Day 12+

Knowledge Base Maintenance

Tier system (pinned/hot/warm/cold), automatic weekly pruning, canonical URL deduplication.

Ongoing

Self-Improvement

SAI reads papers, proposes AGENTS.md changes, human approves each diff. Each approved change is traceable to the paper that inspired it.

SAI is what happens when you stop treating an AI assistant as a static tool and start treating it as an evolving system. The key insight: let it propose changes to itself, but never let it apply them alone.