Here's what I've been building lately. Each project started with a problem that bugged me or a "what if" moment that wouldn't leave me alone. Some succeeded beyond my wildest expectations, others taught me valuable lessons the hard way. All of them shaped how I think about code, AI, and what's possible when you're willing to experiment.


Curious About Any of These?

Each project has its own story—the late nights debugging, the eureka moments, the pivots that saved everything. I love talking through the technical choices, the mistakes I made, and what I’d do differently.

Let’s chat about it →

🔬 Current Rabbit Hole: Long-Term AI Coherence Study

What I'm obsessing over right now: I stumbled across Backlund & Petersson's "Vending-Bench" paper (arXiv:2502.15840) and it perfectly captured something I'd been noticing in my own agent work. Why do LLMs crush individual tasks but completely lose the plot in long-term scenarios? I'm replicating their research because I need to understand this limitation—it's the difference between a clever demo and something you'd actually trust to run your business.

Currently Deep Diving Nights & weekends project
Current Research AI Coherence Autonomous Agents
The Paper That Started It All i

The Digital Office Experiment

I got tired of managing repetitive tasks and wondered: what if I could clone myself digitally? Not just one AI assistant, but a whole team of specialized agents that could actually talk to each other, negotiate, and evolve. I spent months with Google ADK building custom MCPs for ESPO CRM and Zoho, creating this weird digital ecosystem where agents modify each other's instructions based on what they learn. The breakthrough came when I realized they needed memory and personality, not just function calls. Now I have what feels like a virtual office that actually gets smarter over time.

Freed up 70% of my routine work Response time went from hours to seconds
Multi-agent Personal Productivity Learning Systems
i

The Meta-Agent That Builds Agents

Building agents was taking forever—endless iterations, debugging conversations, testing edge cases. So I built an agent that builds other agents. It listens to what you want, builds it, tests it, breaks it, fixes it, and does it all again until it works. The weird part? It debugs by talking to itself about what went wrong. Took weeks → hours off my development cycle.

95% fewer development iterations Weeks of work → afternoon projects
Meta-Development Self-Debugging ADK Integration
See It Build Itself

My Answer to Expensive Market Research

Small businesses can't afford the market intelligence that big companies take for granted. I built this because I was tired of making decisions in the dark. It pulls data from everywhere, connects dots I'd never see manually, and actually predicts trends with scary accuracy. Started using it for my own projects, now friends ask me to run analyses for them.

Tracks 100+ market signals 60% cheaper than traditional research 85% accuracy on 6-month predictions
Data Mining Pattern Recognition Predictive Analysis
i

Content Machine for Wellness+

A wellness startup needed 100+ pages of content weekly. Writing it manually would've killed their budget and my sanity. So I built this Django-powered content engine with GPT-4 that understands wellness trends, personalizes recommendations, and actually writes in a voice that doesn't sound robotic. The engagement numbers surprised everyone, including me.

100+ pages weekly without breaking a sweat Content costs dropped 80% User engagement tripled
Content Automation Personalization Django + GPT-4
i

When AI Started Making Phone Calls

Cold calling is soul-crushing work. I thought: what if AI could handle the grunt work and only pass me the good leads? Built this voice AI system with AutoGen Studio that adapts its pitch based on how the conversation flows. It's weirdly good at reading vocal cues and knows when to pivot. Tripled call capacity while I focused on closing deals instead of dialing numbers.

300% more calls without hiring anyone 90% accuracy in lead qualification
Voice AI Sales Automation AutoGen Studio
i

The Photo Shoot That Never Happened

Product photography was eating budgets alive—every variant, every angle, every seasonal update meant another expensive shoot. Enter Flair.ai. Now I generate product photos that look professionally shot, marketing creatives that adapt to trends, and branded content at whatever scale you need. The time savings alone changed how fast we could launch campaigns.

Production time cut by 90% 70% cost savings vs traditional photography 5x faster campaign launches
AI Photography Creative Automation Flair.ai Integration
i

Speaking Every Language Without Learning Any

Client needed training videos in 8 languages. Traditional approach: hire 8 presenters, 8 sets of equipment, 8 editing cycles. My approach: one recording session with HeyGen, instant multilingual avatars that look and sound natural. When they want updates, I change the script and regenerate. No reshoot, no drama, no budget explosion.

80% less production time 60% savings on localization
AI Avatars Multilingual Content HeyGen Magic
i

Personalized Videos at Industrial Scale

Mass personalization used to be an oxymoron. I built this workflow that takes one template and generates thousands of personalized videos—different names, different data, different languages, all looking like they were made specifically for each viewer. The secret sauce is in how the AI presenter adapts tone and pacing based on the content.

Instant personalization at any scale Supports 40+ languages naturally
Mass Personalization Video Automation AI Presenters
i

Mapping the Future of Work

What happens to jobs when AI gets really good? I dove deep into workforce data from 2000-2025, collaborating with Google Deep Research to understand how roles evolve, which skills become valuable, and what new jobs emerge. The patterns that emerged were fascinating—and not what most people expect. Turned the research into an interactive visualization because spreadsheets don't tell stories.

Labor Economics Future of Work Data Visualization Google Research
Explore the Data

Who's Really Leading AI Automation?

Everyone's talking about AI, but who's actually worth listening to? I analyzed 100+ AI YouTubers to map the real influencers, themes, and strategic insights in automation and agentic systems. The results surprised me—the biggest names aren't always the most insightful. Built an interactive explorer so you can discover the hidden gems yourself.

100+ creators analyzed 5 key themes discovered
Influence Mapping AI Community Analysis Interactive Data
Explore the AI Landscape

Web Development Adventures

Before I fell down the AI rabbit hole, I was building web applications the traditional way. These projects taught me that good software is about solving real problems elegantly, not showing off with the latest framework. Built with Django, Wagtail, Tailwind, HTMX, and lots of coffee.

The PDF Processor That Never Sleeps

Documents are the bane of business existence. I built this processing platform because watching people manually convert PDFs for hours was painful. It handles massive volumes, keeps everything secure, and has the kind of uptime that lets you forget it exists—which is the point.

99.9% uptime (yes, really) Enterprise-grade security
Django HTMX Docker

Southeast Asian tax systems are beautifully complex. Each country has its own rules, currencies fluctuate constantly, and regulations change without warning. Built this platform to make sense of it all—real-time calculations, compliance tracking, and enough automation to keep accountants sane.

Covers 5+ countries seamlessly Real-time compliance checks
Wagtail AlpineJS Tailwind

I love Muay Thai, but I hate unfair matchmaking. Too many fighters get matched based on politics instead of skill, and it's destroying the sport's integrity. So I built FightLink—a completely free platform that uses data-driven algorithms to ensure fair fights. It's my love letter to the Muay Thai community and my small rebellion against corruption in combat sports.

100% Free Forever
Passion Project Community-Driven Fair Play Django
Visit FightLink

Selling Paradise in Three Languages

Thailand's luxury property market is wild—international buyers, complex regulations, and language barriers everywhere. Built this platform to handle automated sales workflows, virtual tours, and intelligent lead routing. The multilingual aspect was trickier than expected, but seeing deals close because of smooth user experience made it worth it.

Handles 3 languages fluently Fully automated workflows
Wagtail AlpineJS Docker

Digital Avatars That Don't Suck

Most virtual avatars look like they escaped from 2005. I wanted to build something that felt real—AI-driven personalization, real-time rendering, and enough technical sophistication to pass the uncanny valley test. The WebGL integration was a nightmare, but the final result actually looks human.

Custom AI models Real-time rendering
Django WebGL TensorFlow

Wellness Tourism Made Simple

The wellness industry loves complexity—multiple providers, different booking systems, payment headaches. I streamlined the entire experience into one platform. Multi-language support for international clients, integrated payments that actually work, and booking flows that don't confuse people.

Multiple providers, one platform Payments that just work
Django HTMX Stripe

E-commerce for a Local Artist

A local creator was drowning in manual order management. Built them a custom e-commerce platform that handles everything automatically—inventory tracking, order processing, customer communications. Mobile-first design because that's how their customers shop. Now they can focus on creating instead of spreadsheets.

100% automated operations Mobile-first design
Wagtail Tailwind Stripe

Want to know more about any of these? I love talking shop about the technical details, the challenges I faced, and what I learned. Drop me a line →