Forem: Andrew Gordon Browne

🩺Rapid Screening for Cardiovascular Risk: A High-Sensitivity Triage Approach

Andrew Gordon Browne — Mon, 09 Feb 2026 18:38:56 +0000

Cardiovascular disease remains a leading cause of global mortality. However, early detection often relies on subjective surveys or time-intensive lab results. In high-volume clinical environments or low-resource settings, every second counts.

For my submission to Byte 2 Beat: Hack4Health, I developed the Rapid Risk Indicator (RRI) set—seven verifiable "bytes" of health data that can be collected in just a few minutes to provide immediate clinical triage.

1. Problem Framing

The goal was to bridge the gap between complex diagnostics and rapid screening. By stripping away subjective "soft" indicators, I focused on a "high-signal" approach that uses only objective, non-invasive data.

2. Methods & Data Strategy

I utilized the CDC’s Behavioral Risk Factor Surveillance System (BRFSS) dataset, which provides a robust foundation of over 250,000 records.

The "7-Byte" Feature Set

To ensure clinical reliability, the model uses only these seven indicators:

Demographics: Age, Sex
Vascular History: Prior Stroke History
Biometrics: BMI
Lifestyle: Smoking Status, Alcohol Consumption
Systemic Access: Healthcare Access

Modeling Strategy

I employed a logistic regression model with a 60/20/20 train-validate-test split.

Class Imbalance: Given the 10% prevalence of heart disease in the dataset, I utilized balanced class weighting.
Optimization: This strategy specifically optimizes for recall, ensuring that the cost of a missed diagnosis (False Negative) is minimized in a triage context.

3. Statistical Validation

All seven indicators achieved a p-value of < 0.001, confirming their validity as significant predictors. Age emerged as the primary driver of cardiovascular risk with an Odds Ratio of 2.59.

4. Results & Evaluation

On the unseen test set (n = 50,736), the framework demonstrated its utility as a powerful clinical safety net:

Sensitivity (Recall): 73% – The model successfully flagged nearly 3 out of every 4 cardiovascular events using only 7 non-invasive questions.
Negative Predictive Value (NPV): 96% – If the model classifies a patient as low-risk, there is a 96% probability that they are healthy, providing high confidence for clinical discharge.

Calibration: How Reliable are the Risks?

Calibration curve analysis showed that while the model is a powerful tool for ranking and triaging risk, it remains conservative in its absolute probability estimates. This confirms its utility for triage ranking rather than absolute diagnosis.

The calibration curve compares predicted probabilities against actual cardiac event proportions.

5. Conclusion

The Rapid Risk Indicator framework demonstrates that a minimal "byte" of objective data can provide a high-sensitivity safety net for heart disease detection. By focusing on p-verified hard indicators and high recall, we can provide a tool that is fast, scientifically ironclad, and ready for real-world clinical implementation.

Building a "Chief of Staff" with Gemini 3: High-Reasoning Agents Beyond the Chatbot

Andrew Gordon Browne — Thu, 29 Jan 2026 17:50:51 +0000

The Problem: AI Noise

Executives don’t need more chatbots. They need high-fidelity intelligence. Most LLM applications today focus on "chatting with data," but for a C-Suite leader, the conversation is often friction. They want the answer, the ROI impact, and the proof.

That’s why I built Lighthouse 3 for the Gemini 3 Global Hackathon.

The Architecture: High-Reasoning & Grounding

Lighthouse 3 isn't just a wrapper; it’s an autonomous research agent built on a "Lean Executive" architecture.

1. The Brain: Gemini 3 Pro

I utilized Gemini 3 Pro Preview with the thinking_level=HIGH configuration. This was the "secret sauce." It allows the model to perform multi-step reasoning to find "Hidden Connections"—linking disparate events like infrastructure pivots to secondary energy market shifts.

2. The Grounding Engine

Using the google-genai SDK, I integrated Google Search Grounding. This ensures the agent isn't hallucinating on old training data, but reacting to market shifts from this morning.

3. The "Thought Signature"

To solve the "Black Box" problem, I implemented a Thought Signature. Every report includes a transparent log of the agent's internal reasoning path so executives can trust how the AI arrived at its strategic advice.

🛠 The Engineering Hurdle: Cloud Run & The Missing Files

Infrastructure is just as important as model logic. I deployed the portal using Docker on Google Cloud Run, but I quickly hit a fascinating hurdle: Deployment Data Persistence.

In the serverless environment of Cloud Run, my generated Markdown briefings were initially "vanishing" due to container lifecycle resets and deployment sync issues.

The Fix:

Absolute Pathing: I moved away from relative paths that behaved inconsistently within the container environment.
The .gcloudignore Precision: I engineered a custom .gcloudignore file to filter out virtual environment bloat while "force-allowing" the mission-critical /reports directory.

"In the cloud, your infrastructure defines the reliability of your intelligence."

Watch the Demo

What’s Next?

Lighthouse 3 is just the beginning. The roadmap includes:

Dynamic Persistence: Moving from the container filesystem to Google Cloud Storage (GCS) for instant, "no-deploy" updates.
Enterprise Integration: Direct delivery to encrypted executive Slack channels.
Long-Term Memory: Tracking strategic predictions over time to create a feedback loop for the model's own reasoning.

The Accidental Portfolio: A Hobbyist’s Guide to Useful Things and AI Second Opinions

Andrew Gordon Browne — Tue, 20 Jan 2026 18:55:27 +0000

This is a submission for the New Year, New You Portfolio Challenge Presented by Google AI

About Me

I am a highly motivated "hobby developer of useful things." To me, engineering is a playground, not a cubicle. I specialize in projects that start as a curious "what if?" at 2:00 AM and somehow survive long enough to be hosted on the internet. I build because I love the puzzle, and I keep it humble because I’ve seen enough "undefined is not a function" errors to know my place in the universe. My portfolio, "The Accidental Portfolio," is a curated set of these survivors.

Portfolio

How I Built It

This site is a "Static-First" labor of love, built with Next.js 16 (using Turbopack) and TypeScript.

I took a somewhat stubborn approach: I did all the heavy lifting manually (Content, JSON, Markdown) to ensure the foundation was solid. I then integrated the Gemini 3 Flash API to act as a "second opinion."

The AI doesn't run the show here; it’s more like a technical co-pilot that looks at my work and offers "Hobbyist Insights"—essentially explaining what I was trying to do and spotting the logic I may have accidentally misplaced during a late-night coding session. I used Google AI Studio to fine-tune these prompts so the AI would sound less like a robot and more like a peer who’s seen my browser's search history.

What I'm Most Proud Of

Honestly? I’m proud that it’s actually live.

On a technical level, I’m most proud of the Dockerized Deployment. It was a humbling reminder of the "it works on my machine" curse. I spent quite a bit of time troubleshooting why the cloud version was empty, only to realize I hadn't explicitly told my Dockerfile to copy over my static data files. Solving that "missing cargo" puzzle and finally seeing my projects and AI insights render correctly in a live Linux container was my biggest "fist-pump" moment. It’s not just a site; it’s a survivor of my own deployment oversight.

🏰 A Most Illuminating Excursion: The Staff of Westmarch Attend Google's 5-Day AI Agents Intensive

Andrew Gordon Browne — Fri, 05 Dec 2025 17:33:37 +0000

When the invitation arrived for the 5-Day AI Agents Intensive Course with Google, the household staff of the Esteemed House of Westmarch insisted---quite vigorously---that they, not I, ought to attend.

And so, armed with curiosity, polished shoes, and a travel satchel filled with architectural diagrams, they set off together to discover the modern disciplines of autonomy, memory, tooling, evaluation, and multi-agent coordination.

What follows is their report, delivered at the weekly staff assembly.

🎩 Jeeves --- Day 1: "On Proper Agent Conduct & Orchestration"

"Ladies and gentlemen," began Jeeves, straightening his cufflinks,
"Day 1 offered a most civilised taxonomy of agentic behaviour:
planning, action, autonomy, and architectural patterns for multi-agent societies."

He was particularly delighted to discover that the course described
orchestration as a first-class discipline.
"At last," he declared, "a formal endorsement of the butler's role in
coordinating a well-governed estate. One cannot simply allow agents to mill about without supervision."

He now insists that every task in Westmarch begin with a properly
structured plan.

📚 Perkins --- Day 2: "Tools Fit for Scholars, Librarians, and Agents Alike"

Perkins returned from Day 2 in a state of near-euphoria.

"Tools!" he exclaimed, balancing six notebooks on one arm. "External
capabilities! The splendid Model Context Protocol (MCP)! A universal method for agents to discover and wield tools with scholarly elegance!"

According to him, converting Python functions into callable agent tools was "as invigorating as opening a fresh crate of encyclopaedias." He now labels every tool in the Estate with a small brass plate and occasionally bows to them. We humor him.

✒️ Miss Pennington --- Day 3: "On Memory, Sessions, & the Delicate Art of Remembering"

Miss Pennington took copious notes during the sessions on short-term context, long-term memory, and context engineering. She has since reorganized the library, the archives, the attic, and---somehow---the cellar.

"Sessions," she explained, "are the parlour conversations of the moment; memory is the grand ledger spanning generations."

Inspired by the course, she introduced a tagging system so comprehensive that Jeeves fears she may eventually catalog him.

🕯️ Lady Hawthorne --- Day 4: "Evaluations, Propriety, & the Pursuit of Agentic Excellence"

Lady Hawthorne found her true calling on Day 4.

"Evaluation," she announced upon returning, "is the beating heart of quality. One must maintain logs, traces, and metrics." (She now issues gentle reprimands if anyone logs ambiguously.)

She was especially fond of LLM-as-a-Judge, which she described as "a perfectly reasonable system, provided the judge is me."

🎩📚✒️🕯️ The Whole Staff --- Day 5: "A2A Protocol & Life as a Multi-Agent Household"

By Day 5, the staff had mastered Agent2Agent (A2A) communication, the very fabric of multi-agent collaboration.

"We discovered," said Jeeves, "that the Estate already was a
multi-agent system. We merely lacked the formal API."

Since returning, the four of them exchange structured messages, route tasks, evaluate one another's work, and escalate matters with unnerving professionalism.

The house has never run so smoothly.

🎬 The Capstone Project --- A Tour of the Estate

With their studies complete, the staff insisted that the final
assignment of the 5-Day AI Agents Intensive Course with Google deserved a proper Westmarch flourish. The Capstone called for a demonstration of everything they had learned---orchestration, tools, memory, evaluation, and multi-agent coordination---and the household saw only one fitting approach: a guided tour of the Estate itself.

"Education," noted Jeeves, "is most effective when one provides a polished example." Lady Hawthorne approved, provided the lighting was dignified.

The resulting film capped---pun unfortunately intended---our Capstone submission and stands as a concise artifact of the week's lessons.

📺 Project Overview Video (2 minutes)

A brief excursion through the architecture and personalities of the House of Westmarch:

Final Reflections from the Staff

The 5-Day AI Agents Intensive affirmed what Westmarch had long
suspected: Agents flourish when given structure, tools, memory, evaluation, and colleagues.

The staff returned wiser, better coordinated, and---if Lady Hawthorne is to be believed---"finally operating at a standard befitting the Estate."

And Jeeves assures me we are now fully prepared for production
deployment.