Show HN: Swiftward – on-prem policy engine for LLM guardrails and UGC moderation(github.com)
1 pointby joker3d6 hours ago2 comments
  • 6 hours ago
    undefined
  • joker3d6 hours ago
    Hi HN — I'm Konstantin, building Swiftward.

    I've been building distributed backend systems for a while. Working on LLM-based systems recently, I noticed a structural issue: policy logic (validators, guardrails, allow/deny rules) usually lives inline in application code.

    That works at first. But when you need versioning, shadow testing, rollback, or reproducible decisions — it gets awkward fast. Especially on-prem.

    Swiftward is my attempt to extract that logic into a dedicated policy runtime.

    It's a self-hosted engine that evaluates events (LLM prompts/outputs, user actions, internal signals) against versioned rules and produces deterministic decisions with full execution traces.

    What it focuses on:

    - Deterministic evaluation (same input + version = same result)

    - Versioned rules with shadow runs and rollback

    - Full decision trace for debugging and audit

    - Postgres-first, fully on-prem deployment

    The repo has Docker Compose demos with pre-built images you can try locally. The engine itself is commercial (self-hosted license) — looking for design partners and paid pilots.

    I'm new to Trust & Safety as a domain and looking to validate where this is useful and where it breaks.

    Website + docs: https://swiftward.dev

    Repo + demos: https://github.com/disciplinedware/swiftward

    Happy to answer technical questions.

    • joker3d6 hours ago
      Here's what a simple policy rule looks like:

          rules:
      contains_hello:
        all:
          - path: "event.type"
            op: eq
            value: "create_post"
          - path: "signals.text_contains_hello"
            op: eq
            value: true
        effects:
          verdict: rejected
          response:
            blocked: true
            reason: "Post contains 'hello'"
      Every decision is stored in Postgres:

          {
      "entity_id": "user_2558",
      "decisions": {
        "matched_rules": ["contains_hello"]
      },
      "effects": {
        "verdict": "rejected",
        "response": { "blocked": true, "reason": "Post contains 'hello'" }
      }
      ...
    }
      To try locally (~2 min):

          git clone https://github.com/disciplinedware/swiftward
    cd examples/demo-minimal
    docker compose up
      No UI in minimal demo — just logs. There's also a UGC demo with full UI, A/B testing, and LLM classification, but it downloads ML models on first run (~10-15 min).