Skip to content
RJ
All work

Case study

LocalOps AI

Private / local AI engineering prototype

A local-LLM decision prototype that grounds an incident against policy, returns a structured decision and hands the final action back to deterministic Python logic.

Status
Technical prototype
Focus
Local AI · grounding · Python orchestration

Technical summary

How the prototype combines a local model, policy context and application logic.

Local inference
Qwen via LM Studio
Grounding
Internal policy document
Application
Python
Output
Structured JSON
Action
P1 escalation record

01 — The question

Can a local model reason over policy without owning the action?

Many operational AI ideas become risky when the model is allowed to improvise both the interpretation and the action. LocalOps AI separates those responsibilities.

The local model reads the incident together with the supplied policy context. Python then consumes a structured result and applies explicit rules around what happens next.

02 — Architecture

LLM for interpretation. Code for control.

Local decision pipeline

Model reasoning is bounded by policy context and application logic.

  1. 01

    Incident input

    A fictional incident is submitted to the Python application.

  2. 02

    Policy context

    The relevant incident policy is supplied as grounding context.

  3. 03

    Local LLM

    The model interprets the incident against that supplied policy.

  4. 04

    Structured decision

    The response is returned in a predictable structured format.

  5. 05

    Deterministic logic

    Python checks the structured result and owns the action logic.

  6. 06

    Action or escalation

    Critical paths can be escalated rather than left to free-form model output.

03 — Prototype in action

A fictional P1 incident, interpreted and routed through escalation logic.

The model supplies a structured classification; the application decides whether to create an escalation record.

01

Incident input

"The production database is completely down and customers cannot complete transactions."

02

Local Qwen decision

{
  "priority": "P1",
  "notify": [
    "Incident Manager",
    "Database Operations team"
  ],
  "assign_to": [
    "Database Operations team"
  ],
  "acknowledgement_minutes": 10
}
03

Python-controlled action

  • JSON parsed by Python
  • Priority evaluated as P1
  • P1 escalation rule triggered
  • Escalation record created
  • Status: pending_escalation
Incident + policyQwen interpretsstructured JSONPython validatesescalation logic

AI responsibility

Reads the incident against the supplied policy and returns the required structured classification.

Application responsibility

Parses the JSON, checks the priority and controls whether an escalation record is created.

04 — Privacy & proof

Safe portfolio evidence, not company data.

The case study uses a fictional company, fictional incidents and demonstration policy data. No employer names, hostnames, addresses, real internal policies or production records are shown.

The portfolio demonstrates the architecture and reasoning boundary, not confidential operational data.

05 — What it demonstrates

Private AI, Python and operational thinking in one workflow.

The prototype shows how I think about AI inside operational systems: use the model where interpretation helps, constrain its context, make its output machine-readable and keep important actions in explicit application logic.

Built by Roselyne June

View all work →