status: prototype completed environment: local deployment · 8GB RAM type: personal R&D case study
project_showcase / genai_ops

AI-powered operations intelligence assistant

A local Generative AI assistant built to explore how prompt-engineered LLM workflows could support operational risk analysis and executive-style reporting — applying 20+ years of airline operations context to a self-hosted AI stack.

Ollama Docker Open WebUI Prompt engineering Local LLM deployment
local intelligence pipeline ● no cloud calls
Operational dataincident logs + KPI notes
Ollama runtimeself-hosted local LLM
Prompt workflowsrisk + KPI + summary
Management outputstructured draft summary
data remains on-device
01 · objective

Why this project exists

Most operational risk reporting in airline environments is manual: someone reads incident logs, performance data, and audit notes, then writes a summary for leadership. This project explored whether a locally-run, open-source LLM could assist with that workflow — surfacing recurring risk patterns and drafting executive-style summaries from operational data, entirely on local hardware with no cloud API calls.

02 · architecture

How it was built

A fully local pipeline with no external inference or cloud API calls. The language model ran locally through Ollama on a single machine, while Open WebUI provided the Docker-hosted interface.

01
Data input
Aviation safety & operational datasets, structured as text for prompting
02
Ollama
Local LLM runtime serving the model on-device
03
Prompt workflows
Structured prompts for risk analysis, KPI assessment, summarization
04
Open WebUI
Docker-hosted chat interface for querying the assistant
03 · in practice

Illustrative example

The original session logs no longer exist — the project was decommissioned after running into hardware limits (see note below). The exchange below illustrates the kind of prompt workflow the assistant was designed to support, reconstructed for this showcase rather than copied from a saved transcript.

open-webui — local session
> Summarize recurring turnaround delay causes from the last 30 incident logs, in 3 bullet points for a station manager briefing. Recurring causes identified: • Late inbound connections accounted for the largest share of delays, concentrated on two routes • Ground equipment availability caused secondary delays during peak turnaround windows • Documentation handoff between shifts added avoidable buffer time on ~15% of flights Suggested focus: equipment scheduling during peak windows and shift handoff standardization.
operations intelligence summary illustrative prototype output
04 · business context

Where this prototype fits

The technical prototype was designed around a real operational reporting problem: turning fragmented incident information into a concise management view without removing human review or operational accountability.

Who it is for

Station operations managers, network controllers, and safety or compliance teams that currently prepare shift-end, weekly, or incident-pattern summaries manually.

The manual workflow addressed

Based on my operational experience, one shift-end or weekly incident summary can require 30–60 minutes of log review and writing, depending on incident volume. This is an experience-based estimate, not a measured project result.

Path to production

A team-ready version would require a quantized or appropriately hosted model, a proper data pipeline, access controls, output validation, and testing with real operational users before any leadership reporting use.

05 · what it demonstrated

Outcomes

prototype demonstrated
A fully local LLM pipeline can run structured operational-analysis prompts without any cloud dependency or data leaving the device.
prototype demonstrated
Prompt-engineering patterns for summarization and risk-pattern extraction translate well from general LLM use to operations-specific reporting tasks.
constraint hit
8GB RAM was the practical ceiling for running a local model alongside Docker and Open WebUI at usable response speed.
next step identified
A quantized smaller model or an appropriately configured hosted environment could reduce the local hardware constraint while retaining the same prompt-engineering approach.
note — this prototype ran locally on an 8GB RAM machine and was later decommissioned to free system resources. The original screenshots and session logs are no longer available. The prompt exchange and dashboard above are clearly labelled reconstructions that illustrate the intended workflow rather than captured production evidence.
06 · stack

Tools used

OLOllama DKDocker OWOpen WebUI AIGenerative AI / LLMs PEPrompt engineering