Hackathon · Ironhack Madrid

Margaret Hamilton, restored black-and-white portrait — Margaret Hamilton, 1969 · Draper Laboratory; restored by Adam Cuerden. Source and reuse information · Wikimedia Commons

Margaret Hamilton — Ticket agent

An AI agent that investigates data tickets on Databricks and proposes the fix. Built in two days for the Alenia × Ayvens AI Ticketing Challenge.

Date: 15–16 Oct 2026
My role: Team lead · integration
Delivered: Databricks App, 236 tests

OpenProject Databricks LangChain Llama 3 Streamlit Python

01 · Challenge

Resolve fleet-data tickets

Ayvens IT teams handle thousands of support tickets a month about vehicle-fleet data. Alenia asked for an AI workflow that understands a ticket, investigates the data and proposes a resolution.

UnderstandInvestigateResolve

Alenia challenge brief describing the three stages Understand, Investigate and Resolve — Challenge brief · Alenia

OpenProject Databricks LangChain Llama 3 Streamlit Python

02 · Platform

The challenge platform

Alenia provided OpenProject for tickets, Databricks Serverless for data and compute, Vector Search and a shared LLM endpoint. Any change had to pass human approval.

Architecture diagram connecting OpenProject, Databricks Serverless, Vector Search, the agent layer and human approval — Platform architecture · Alenia

OpenProject Databricks LangChain Llama 3 Streamlit Python

03 · Solution

What we built

One agent that reads the ticket queue, routes each ticket, checks the data and waits for approval before it writes anything.

01 · Intake OpenProject queue Open tickets, in priority order.
02 · Route Router Keyword rules first; the LLM only if none match.
03 · Use case UC1 · UC2 · UC3 Normalization, reconciliation, data verification.
04 · Evidence bronze · silver · gold SELECT queries on the relevant layers.
05 · Output Proposal The fix, with the evidence behind it.
06 · Control Human approval A person approves or rejects.
07 · Apply Write back Ticket status and comment, a solution row, and an email to the delegates concerned.

OpenProject Databricks LangChain Llama 3 Streamlit Python

04 · Agents

Three use cases

Each use case is a separate agent that checks one kind of data problem.

UC1

Normalization

A local code that is unmapped, or mapped to the wrong value.

Anahi · agent 1

UC2

Monthly reconciliation

The gold Fleet Report against the silver and bronze layers.

Karina · agent 2

UC3

Data verification

A contract or vehicle missing or wrong in bronze or silver.

Karina · agent 3

OpenProject Databricks LangChain Llama 3 Streamlit Python

05 · Human in the loop

Approval before any write

Processing tickets only runs SELECT queries. Nothing is written to OpenProject or Databricks until a person clicks Approve.

Process: read-onlyApprove: write

Margaret Hamilton app showing an OpenProject ticket queue and a note that processing runs read-only SELECT queries until approval — Margaret Hamilton app · ticket queue

OpenProject Databricks LangChain Llama 3 Streamlit Python

06 · Results

Results per ticket

For every ticket the app shows the evidence found, the proposed fix and the solution row it recorded, including who gets the email.

EvidenceProposalSolution row

Synthetic ticket and company data. Email addresses redacted.

Margaret Hamilton app showing step-by-step results, a summary and a solutions table with recipient addresses redacted — Margaret Hamilton app · step-by-step results

OpenProject Databricks LangChain Llama 3 Streamlit Python

07 · Deployment

Deployed on Databricks

The Streamlit app runs as a Databricks App. Merging to main deploys it to production.

236 automated tests108 commits in 3 days

Workspace, URL, creator and app ID are redacted.

Redacted Databricks Apps deployment page for the Margaret Hamilton app — Margaret Hamilton app · Databricks deployment

OpenProject Databricks LangChain Llama 3 Streamlit Python

08 · Cost

Cost per ticket

Estimated with the challenge cost model, not measured in production.

Estimated cost and handling time per ticket
Measure	Ticket agent	Analyst L1
Cost per ticket	$0.0009	$6.39
Time per ticket	30 s	15 min

1,000 tickets per month

≈ $6,400 saved per month

242 h analyst time saved per month

Analyst L1 salary: $45,000 per year.
Working time: 1,760 hours per year, or $25.57 per hour.
Analyst handling time: 15 minutes per ticket.
LLM price: $0.90 per 1 million tokens.

OpenProject Databricks LangChain Llama 3 Streamlit Python

09 · Model

Model cost and latency

The agent runs on Llama 3.3 70B. Figures from the challenge's model-cost data.

Estimated cost for 1,000 tickets and model latency
Model	Cost / 1K tickets	Latency
gemma-3-12b	$0.1363	1.16 s
llama-4-maverick	$0.4143	0.63 s
databricks-meta-llama-3-3-70b-instruct Current	$0.7830	0.84 s
gpt-5-mini	$1.73	1.03 s
claude-haiku-5-5	$5.18	1.99 s

OpenProject Databricks LangChain Llama 3 Streamlit Python

10 · Team

Team and roles

Four people. I led the team and built most of the solution.

Team lead · integration

Sebastian

Led the team, built most of the solution, connected the Databricks APIs and OpenProject tokens, and polished the result.

Agent 1 · structure

Anahi

Created agent 1 and proposed the solution structure.

Agents 2 & 3

Karina

Created agents 2 and 3.

Verification

Umit

Ran the tests to verify that the solution worked.

Pi + gentle-ai Configuration layer

Claude 5.5 Orchestrator

DeepSeek-4 Flash Code

Gemini Verification and tests

OpenProject Databricks LangChain Llama 3 Streamlit Python

Margaret Hamilton · Ticket agent

From ticket to approved fix.

Alenia × Ayvens AI Ticketing Challenge · Ironhack · Madrid, 15–16 Oct 2026

Back to the portfolio

OpenProject Databricks LangChain Llama 3 Streamlit Python