Osistan

AI operations assistant that actually runs in production.

Problem

Operations teams drown in repetitive queries, email triage, deadline tracking, and knowledge that lives in people's heads. Commercial AI tools either cost too much or don't understand your specific business context.

Approach

3-layer architecture — deterministic engine handles cron jobs, email parsing, deadline math with no LLM involvement and no hallucination. Context assembler builds optimal token windows with position-aware placement. Intelligence layer only thinks — never fetches data or parses files.

Started with 50 tools. The LLM was spending its intelligence on tool selection instead of thinking. Reduced to 3 write channels: db_write, slack_send, file_create. Everything changed.

Key Details

• Runs on Raspberry Pi 5 (8GB RAM) — 24/7 production
• 39K email chunks embedded with pgvector (3.4ms search)
• Event sourcing for task management
• Karpathy wiki pattern — knowledge compiled once, kept current
• Self-observation — tracks its own action outcomes over time
• Progressive soul file — 27K tokens → 3.9K core + on-demand sections
• Cost: ~$40/month serving a team of 8

Challenges

Token budget management. The full soul file was 27K tokens — too expensive for every request. Solution: progressive disclosure. A 3.9K core personality loads always. Domain-specific sections (contracts, escalation, reporting) load on demand based on the conversation topic.

Email noise. 472 email senders across 5 categories. Without filtering, the model would mix up bus company route emails with partner communications. Solution: sender tagging with firm-specific filtering. When someone asks about a specific company, only that company's emails surface.

Self-observation without feedback loops. The agent tracks whether its tasks get cancelled within 24 hours (bad signal) or ignored for 48 hours (noise signal). This data feeds monthly reviews that refine the soul file — but the agent never modifies its own instructions automatically. A human reviews the patterns first.

Stack

Claude Opus 4.6 + Sonnet 4.6 · Python · PostgreSQL · pgvector · Slack Bolt · Raspberry Pi · Tailscale · PM2