Alex NovaraExperience
Software engineer building reliable web platforms and practical AI tools
11 years of professional experience
PDF updated
Leads the platform team of a fictional digital publisher: the reading app, the editorial tools and, since 2024, the AI features of both.
- Own the architecture of the reading platform and its editorial back office
- Introduce retrieval-based search and summaries with an evaluation suite
- Mentor four engineers and run the weekly design review
- TypeScript
- Next.js
- PostgreSQL
- pgvector
- Python
- LangGraph
Reading assistant
March 1st, 2024 – present
Answers readers' questions from the fictional publisher's own archive, with sources.
- Problem
Readers searched an archive of twenty years by keywords and rarely found the article they remembered.
- My role
Technical lead
- Solution
Designed retrieval over the archive with pgvector, a small answering agent and an evaluation suite of two hundred real-looking questions.
Context
The fictional Lumen Press wanted readers to find old articles by asking, not by guessing keywords.
Approach
We wrote the evaluation questions first, then tried three chunking strategies against them. The agent may only answer with a cited article and says so when it cannot.
Outcome
The assistant shipped to a test group; the evaluation suite now guards every change to prompts, chunking and models.
- Result
Four in five test questions are answered with a correct source; the suite runs on every change.
- Python
- LangGraph
- pgvector
- PostgreSQL
Editorial workbench
June 1st, 2022 – September 30th, 2023
One back office for planning, editing and publishing the fictional publisher's stories.
- Problem
Editors juggled three tools and a shared spreadsheet to publish one story.
- My role
Lead engineer
- Solution
Built a Next.js workbench over one PostgreSQL schema, with drafts, review and scheduled publishing.
Context
A fictional newsroom with three tools and one spreadsheet.
Approach
Interviews with six editors, a shared data model first and a workbench released one desk at a time.
Outcome
The old tools were switched off after four months.
- Result
Publishing a story takes one tool and half the steps.
- TypeScript
- Next.js
- PostgreSQL
Headline tester
February 1st, 2023 – May 31st, 2023
A small tool that compares headline variants on a share of the fictional readership.
- Problem
Headline choices were made by gut feeling.
- Solution
Added variant assignment at the edge and a daily report.
- Result
Editors test two headlines a day on average.
- TypeScript
- Redis
“Alex wrote the test questions before writing the assistant. That is why we trusted it with our archive.”

Mira Example, Editor-in-chief, Lumen Press (fictional) A fictional one-person studio for architecture reviews and small AI prototypes for local businesses.
- Review web architectures
- Prototype assistants over company documents
- TypeScript
- Python
- Model Context Protocol
- Docker
Document assistant prototype
September 1st, 2024 – November 30th, 2024
A prototype assistant that answers staff questions from a fictional firm's handbooks.
- Problem
Staff of a small accounting firm asked the same handbook questions every week.
- Solution
Built a prototype exposing the handbooks to an assistant through an MCP server, with a short evaluation set.
- Result
The firm kept the prototype for internal use.
- Python
- Model Context Protocol
- Docker
Architecture review kit
June 1st, 2020 – present
A repeatable checklist and report template for reviewing small web platforms.
- Problem
Each review started from a blank page.
- Solution
Collected the questions of earlier reviews into a checklist and a report template.
- Result
A review takes two days instead of five.
- TypeScript
“Four weeks later our staff stopped asking the same handbook questions.”
Dana Placeholder, Managing partner of a fictional accounting firm Co-maintains a fictional open-source catalogue for small community libraries.
- TypeScript
- React
- Playwright
Community catalogue
March 1st, 2018 – present
An open-source catalogue for fictional community libraries.
- Problem
Small libraries kept their catalogue in spreadsheets.
- Solution
Maintains the React front end and its Playwright suite.
- Result
Used by a dozen fictional libraries.
- TypeScript
- React
- Playwright
Built the online store and the warehouse tooling of a fictional retail group with forty shops.
- Move the store from a monolith to a TypeScript API with a React front end
- Build the stock and picking tools used in two warehouses
- Keep checkout available during seasonal peaks
- TypeScript
- Node.js
- React
- PostgreSQL
- Redis
- GraphQL
- Terraform
Store replatforming
April 1st, 2019 – February 28th, 2021
Moved the fictional retailer's online store to a TypeScript API and a React front end.
- Problem
Every change to the old store risked the checkout.
- Solution
Strangled the monolith route by route behind a GraphQL API, with contract tests for each route.
- Result
Releases went from monthly to daily without a checkout outage.
- TypeScript
- Node.js
- React
- GraphQL
- PostgreSQL
Warehouse picking
March 1st, 2021 – March 31st, 2022
Handheld picking lists for the fictional retailer's two warehouses.
- Problem
Pickers walked the warehouse in the order orders arrived.
- Solution
Built a picking service that groups orders by aisle and a small handheld front end.
- Result
Pickers walk about a third less per shift.
- TypeScript
- Node.js
- Redis
“We replaced the whole store without one checkout outage in a peak season.”
Jonas Sample, Head of e-commerce, Harbor Retail Group (fictional) Started a career on the booking portal of a fictional regional freight company.
- Maintain and extend the customer booking portal
- Write the integrations with the carriers' tracking feeds
- PHP
- Symfony
- PostgreSQL
- Bash
Booking portal
January 1st, 2016 – December 31st, 2018
The fictional freight company's portal for booking and tracking shipments.
- Problem
Customers booked shipments by e-mail and phoned to track them.
- Solution
Extended the Symfony portal with online booking and live tracking from carrier feeds.
- Result
Most bookings arrive through the portal.
- PHP
- Symfony
- PostgreSQL
“A junior who asked the right questions and shipped the tracking integration on time.”

Hana Fictive, Team lead, Northwind Freight (fictional)
Independent R&D
Unreleased research and experiments, apart from the career above. Each says how far it went.
Citation checker
BuiltImplemented and working
Built: a second model pass that refuses answers whose citations do not support them.
- Grounded answers
- Self-checking
- Refusal on missing evidence
Chunking benchmark
TestedTried hands-on
Tested: three chunking strategies compared on the same fictional question set.
- Chunk size
- Overlap
- Retrieval recall
Tool permission model
DesignedUnimplemented design, planned or architected only
Designed: a permission model for agents that act through tools, with read and write scopes.
- Least privilege
- Human approval
- Audit trail
Evaluation dashboard
BuiltImplemented and working
Built: a small dashboard that tracks evaluation scores across prompt and model changes.
- Regression tracking
- Pass thresholds
Technologies
Core
- TypeScript
- Node.js
- React
- Next.js
- PostgreSQL
- PHP
AI engineering focus
A current specialisation, not years of core expertise.
- Python
- LangGraph
- pgvector
- Model Context Protocol
- Docker
- Redis
- Symfony
- Terraform
- Playwright
- GraphQL
- Bash
Education
Informatics, master's degree
Harbour City University
September 1st, 2013 – June 30th, 2015School
A fictional programme; thesis on scheduling freight routes.
Applied informatics, bachelor's degree
Harbour City University
September 1st, 2010 – June 30th, 2013School
Cloud solutions architect
Example Cloud Academy
May 14th, 2021Certification
Evaluating language-model applications
Example Online School
January 8th, 2024 – March 1st, 2024Course
First agent in production
February 1st, 2025Milestone
The fictional reading assistant reached its first test group.