How One Company Cracked Indoor GPS + Why 95% of Corporate AI Projects Fail

Cutting edge tech is nice when it works properly

Author

Rares Serban
September 03, 2025

 👋 Introduction

Hey everyone,

Today we will look at a couple of interesting stories I found regarding using AI in WordPress and the recent MIT study showing that 95% of corporate AI pilots fail.

But first, a deep dive into HyperAR, the company that solved indoor location tracking.
Here's how they cracked the code:

🔍 Deep Dive: HyperAR indoor location tracking

How one company solved the “biggest unsolved problem in retail”

Imagine going to a big shopping mall. You see all these stores everywhere, advertisements taking your attention and making you want to buy things you do not want. You just want to get a new backpack, the brand does not matter. But you do not know where to find backpacks exactly, and going through the whole mall could take hours…

One company, HyperAR, sought to solve the yet unsolved problem of reliable indoor navigation. They wanted customers to be able to easily search for an item (ie: backpack) and then get accurate directions on their phone that would guide them around the shopping mall.

Although at first glance this problem sounds simple, it is actually quite complex. Even though indoor maps exist, GPS doesn’t work indoors to pinpoint the exact user location, and other solutions, like using WiFi or Bluetooth signal strength, did not provide accurate enough location data. The issue is that indoor navigation needs to be 2x as accurate as a strong GPS signal. Aisles in a store are usually about 2 meters wide, so if you have a margin of error of 2 meters, the user’s location might be in the wrong aisle entirely.

A lot of studies, from Apple and Google, have been made to solve this problem, but each ran into limitations:

  • Bluetooth signals are noisy, having an accuracy of about 5 meters

  • WiFi seems more promising, but still not accurate enough either

  • Magnetometers (built into phones) use the earth’s magnetic field, but take several minutes of walking to pinpoint the exact location

  • Computer vision, like Google Street view is clunky for the user, having to point their phone’s camera at surroundings

So how do you build a reliable solution that takes minimal amount of time to give you the initial accurate location, while also re-using existing hardware and being scalable?

The final solution involved a combination of data sources.

1. WiFi location

A map of signal strengths can be built initially to pinpoint a rough location of a phone, by measuring the signal between each individual access point.

The company collected WiFi signal data by sending a member of their team to a site to scan the venue using a phone. Then, proprietary algorithms were used track precise motion between each recorded point, in order to collect as much WiFi data as possible.

Eventually, they build a new version which lets staff at the store do the initial setup scanning. Using WiFi gives a 3 meter indoor accuracy, but this is not enough for turn-by-turn navigation.

2. Motion data

To refine the accuracy to 1 meter, motion data from phones was used. To do this, they used a system called SLAM (Simultaneous Localization and Mapping), which uses the accelerometer, gyroscope and camera to understrand precise device motion.

This gives them motion data, but in it's own coordinate space. The job then is to combine this motion data with WiFi triangulation data and get an accurate location of the user once he starts moving. Algorithms then refine the location to a consistent 1 meter accuracy, which can then be used to provide turn by turn navigation in the store.

Combine this with AR, and you have a really friendly customer product.

Going back to our initial story, the user can now get the mall’s app on its phone, open the map and see an approximate location where he is in an aisle. Once he starts walking, he sees exactly in which aisle he is in and can get turn by turn navigation to all the backpack locations in the big shopping mall.

This story was inspired by Andrew Hart’s (founder of HyperAR) X post:

🚀 New Developments: MIT report shows 95% of generative AI pilots at companies are failing

Corporations are having a hard time

If you work in a corporate environment, and you were probably asked to use tools X and Y which advertise themselves as using AI. And you probably do not want to use such a tool, since it doesn’t help your day to day life at all.

Well, an MIT study called The GenAI Divide: State of AI showed just that: despite $30-40 billion invested in enterprise AI usage, 95% of corporate AI pilots are failing to deliver any measurable business returns. Who would have thought? 😄 

Study scope

The comprehensive research was led by Aditya Challapally and conducted through multiple data sources: 150 executive interviews, surveys of 350 employees, and analysis of 300 public AI deployments across various industries. This robust methodology provides a definitive snapshot of enterprise AI adoption challenges in 2025.

The study reveals what researchers term the "GenAI Divide" - a sharp contrast between widespread AI experimentation and actual transformation. Whileover 80% of organizations have pilot programs for generative AI tools like ChatGPT and Microsoft Copilot, only 5% successfully scaled these initiatives to production with measurable profit-and-loss impact. Most projects remain trapped in "pilot purgatory," never progressing beyond initial testing phases.

Causes of Failure & Success

The primary culprit is not inadequate AI models but rather a fundamental "learning gap". Unlike consumer tools that excel through flexibility, enterprise AI systems fail because they don't retain feedback, adapt to existing workflows, or improve over time. Generic tools become "brittle, overengineered, or misaligned with actual workflows" when forced into complex organizational environments.

The successful 5% share common characteristics: they partner with specialized vendors rather than building internally, focus on specific pain points, and implement systems that can learn and adapt. Interestingly, companies purchasing AI solutions from vendors succeed 67% of the time, while internal builds succeed only one-third as often.

Meanwhile, a "shadow AI economy" thrives with 90% of workers using personal AI tools like ChatGPT for work tasks, often outperforming official corporate systems due to their responsiveness and flexibility. This highlights the disconnect between top-down corporate AI strategies and actual user needs.

This MIT report serves as a wake-up call for the tech industry, especially for senior engineers, since workers have been sounding the alarm bells for quite some time.

🔥 Battle-Tested Tech: Using AI for WordPress development

43.4% of the Web still uses WordPress

WordPress development has changed a lot during the years. If you didn’t know, WordPress is 22 years old 😯 and is still used by a lot of people. Somehow, it even stayed up to date with the changing times, and now AI is used to help developers and users use it better.

Gone are the days of setting up a new WordPress instance and getting a fresh dashboard with empty content and no design. In the past, you had to manually install a theme, add plugins and create content. This was quite frustrating for new users who didn’t have technical experience.

AI is also no longer restricted to outside of WordPress. There are a lot of plugins that now use AI to simplify initial setup, to allow users more time to focus on the actual content and not configuring a theme or plugin a certain way.

Thanks to AI, building a website has gone from a slow grind to a nifty sprint. Tools like 10Web can whip up fully functional websites—from design to content—in minutes, saving developers (and users) hours of tinkering.

This shift is also shaking up hosting providers, who now offer ready-to-go sites instead of blank slates, cutting down customer frustration and support calls. The future? WordPress becomes a platform not just for building sites but for running and evolving them with AI-powered insights, smarter SEO, and real-time improvements that adapt based on user behavior.

📈 Recent Trend: AI security

Security vulnerabilities in AI-powered tools are emerging faster than fixes

Besides the vibe coding security practices (or lack thereof) that we previously discussed in this newsletter, it seems now that AI browsers can also be used to steal your bank credentials 😯 

Brave team recently conducted a study that showcased how AI browsers can be used to steal sensitive information from users. This flaw affected the new Perplexity Comet browser, but was already fixed.

The issue was in the logic that Comet used to summarize websites for users. When it processed site content, the browser couldn’t tell if the instructions were legitimately made by the user or not. In practice, an attacker could embed malicious prompts in a web page to “trick” the AI browser of performing certain actions.

Using the latest technology is great, but please make sure to also consider the security implications that arise with you giving the browser access to an entire web page.

Also, Cursor strikes again, and drops an entire production database 🙃 .

🏆️ Top GitHub Repo: Claude Flow

🌟7k stars+ | AI-powered development orchestration

This tools uses a master AI process that coordinates and assigns work to specialized worker agents. You give it a task, and the AI will try to solve it by delegating work to other agents.

It can be used in two modes:

  • Swarm, best for quick tasks, uses temporary coordination

  • Hive-Mind, best for complex projects with persistent sessions and resume capability

Built By Ruvnet

Quick Start:

# 1. Install Claude Code globally
npm install -g @anthropic-ai/claude-code

# 2. (Optional) Skip permissions check for faster setup
# Only use if you understand the security implications
claude --dangerously-skip-permissions

# Initialize once per feature/task
npx claude-flow@alpha init --force
npx claude-flow@alpha hive-mind spawn "Implement user authentication" --claude

# Continue working on SAME feature (reuse existing hive)
npx claude-flow@alpha hive-mind status
npx claude-flow@alpha memory query "authentication" --recent
npx claude-flow@alpha swarm "Add password reset functionality" --continue-session

🔄 Tech Updates

🗝️ Legacy Revival

🐦‍⬛ X Hits

  1. Watch out for EVM Proxy attacks if you are a smart contract developer

  2. Is it better writing bad code instead of good code? 🤔 

  3. Explaining sub-agents in Claude code

  4. Are AI hallucinations fixed?

Looking forward to hearing from you in the next weeks!

Till next time,
Rares.

Reply

or to participate.