TOJI TOJI
System Architecture

AI-Agent Project Management

Expert agents collaborate via chat · act on tickets · learn from knowledge

Agent Cycle
Office
Chat
→
Knowledge / HR / Meeting
→
Brain
RAG →
Answer
↓
Ticket
→
PM
→
Assigned
→
Report
→
Chat

Situation A new hire asks in chat about company accounts, benefits, and team policies all at once.

Scenario HR and Knowledge agents pull policies, onboarding docs, and FAQs from Brain, answer in the thread, open tickets when a handoff is needed, reflect assignments and reports, and leave a short summary with links in chat.

Dev
Chat
→
Triage
→
Ticket
→
Dev
→
PR
↓
QA
→
Merge
→
Docs
ingest →
Brain

Situation A developer reports in chat that login sessions sometimes drop.

Scenario Agents turn the conversation into a ticket, ship a fix through PR, then review, QA, and merge. The change is documented and ingested into Brain so later, similar questions cite the past issue and decision.

Why Brain?
Without
Agent → Core → ✕

"I have no context about your project, specs, or past decisions."

With Brain
Agent → Core → Brain → ✓

"Based on your spec v2.3, ticket #142 is blocked by the auth migration agreed on 2025-03-12."

Tri-layer Retrieval

Vector (semantic similarity) + Graph (relationship traversal) + SQL (structural hierarchy) — not just "find similar text"

Structure-aware Chunking

Documents parsed into hierarchy (section → paragraph → sentence); RAG returns the chunk and its parent context — richer, more accurate agent prompts

Stable Knowledge

Idempotent ingestion via stable entity IDs — re-processing a document never creates duplicates or knowledge drift

Self-growing

Every cycle feeds back: merged PRs, meeting summaries, and decisions are ingested — Brain grows smarter with each loop

Architecture
Interface
Relay Chat · Group / DM
→
Core LLM · Reasoning
→
Tracker Tickets · Tasks
Lens Dashboard (opt)
↕
Agents
Dev
Triage intent → ticket
Dev code · PR
QA test · bug
Security scan · CVE
Docs ingest · summary
Office
PM roadmap · priority
Knowledge search · RAG
HR onboard · policy
Report digest · status
Meeting notes → tickets
↕
Foundation
Brain Ingest · RAG
Guardian Auth · SSO
Infra K8s · IaC · Vault
Components
Relay Chat interface — group & DM. Agent entry point for users.
Core LLM reasoning layer. Interprets intent, generates responses, decides next action.
Tracker Ticket & task engine. Source of truth for all work and assignments.
Brain Knowledge engine. Ingest pipeline ingests docs; RAG retrieves relevant context for agents.
Guardian SSO & identity. Manages authentication and relationship-based access control.
Infra Infrastructure orchestration — container platform, IaC, secrets management.
Lens (opt) Dashboard & viewer. Optional visual interface for project & metric overview.
Infra Stack
Docker Small
  • Docker Compose
  • Traefik (reverse proxy)
  • Vault (secrets)
  • cert-manager · Let's Encrypt
  • GitHub Actions (CI/CD)
  • GHCR (registry)
Kubernetes Medium · ★ Recommended
  • Kubernetes (kubeadm)
  • OpenTofu (IaC)
  • Vault + Agent Injector
  • Traefik (Ingress)
  • cert-manager (TLS)
  • CSI Storage (SSD · HDD)
  • Calico (CNI)
  • GHCR · GitHub Actions
  • Prometheus · Loki · Grafana
OpenStack Large
  • All K8s tier items
  • Nova · Neutron · Keystone
  • Cinder (iSCSI block storage)
  • Octavia (LB · Amphora)
  • OpenStack CCM + Cinder CSI
  • kolla-ansible (deploy)
  • Glance (image registry)
Data Layer — all scales
Required PostgreSQL Qdrant · Vector TypeDB · Graph
Recommended Redis Loki · Grafana
Workflow (opt) Temporal n8n Airflow ···
Stream (opt) Redpanda Kafka ···
Required Recommended Optional