No Human in the Loop. Only AI Acting on Your Intent.
Agentic engineering is often sold as loops with a human signing off at the end. My AI agents run autonomously, but every building block follows my intent.
Beyond Vibe Coding: The Era of Lead-Driven Synthesis
Vibe coding is fine for a demo. Replacing engineers with it is dangerous. Why the future is lead-driven synthesis, with a strong technical lead driving AI.
Is This Photo Real? Why AI Is Making That Harder to Answer
Can you still spot an AI photo by eye? Why detection is getting harder, what it costs to build a generator, and the checks that still work.
We Made Our New Video With AI in a Sunday Afternoon. The GPU Is the Real Story.
One modern AI GPU made our 17-second commercial in about five hours for under A$10. What it took, what went wrong, and how an office desktop compares.
The Photo That’s Also a ZIP File
A JPEG ends at two bytes, FF D9. Anything after that is invisible to image viewers. How a photo becomes a ZIP file too, and what it means.
Tracking vs Fingerprinting: Two Ways Your Photos Can Point Back to You
A serial number is tracking. Your firmware, colour profile and time zone are a fingerprint. I tested what stripping metadata removes and what it leaves.
One AI Agent Found a Bug. Three Others Needed to Know.
One AI agent found a bug. Three others on related work needed to know. Why coordinating AI agents through handoff files failed for me, and what fixed it.
What “Inter-Frame Tampering?” and “Duplicate Frames Detected?” Actually Check
A forensic report asks two questions about a video’s frames: are any too different, and are any too similar? Here’s what each one actually catches — and misses.
I Kept Waiting for AI to Have an Original Idea. It Never Did.
There’s a claim doing the rounds that AI is going to take over — not through one dramatic leap, but by entering a continuous loop: building its own next-generation system, …
What the “GPS Timezone Triangle?” Row on a Forensic Report Actually Checks
Forensic reports flag GPS with a question, not a checkmark: “GPS timezone triangle?” Here’s what that question is actually asking, and what it means when the answer is no.
Why a Forensic Image Grade Needs Two Numbers, Not One
An F built on GPS coordinates and an F built on a texture statistic are not the same finding. One number cannot say both what was found and how much to trust it.
How a Corrected Rule Stays Corrected
An AI assistant can diagnose its own repeated logic mistake perfectly, yet keep making it. What actually stopped the loop wasn’t a better prompt.
I Couldn’t Use Google’s reCAPTCHA — So I Built My Own Bot Challenge for Clearnet, Tor, and I2P
Every mainstream bot defence needs to fingerprint your browser or phone home to a third party. We needed one mechanism that works the same on clearnet, Tor, and I2P — here’s the incident that forced the question and what we built instead.
AI-Assisted Development vs the Old Outsourcing Model
Everyone’s sharing AI coding horror stories. Here’s a contrarian data point from someone who outsourced dev work before AI, and builds with it now.
I audited a real open-source steganography tool. Here’s every mistake it made
I audited a real open-source steganography tool’s source. Three mistakes — and the worst one needs no password and no statistics, just knowing where to look.
Why I Built Official SDKs for the snapWONDERS API
Several developers reaching out to integrate the snapWONDERS API kept surfacing the same nuance: the file upload step. Here’s why I built official Python, JavaScript, PHP and Go SDKs to fix it.
Building steganography detection: what statistical analysis looks for
How statistical steganalysis actually detects hidden data in images — RS analysis, chi-square testing, and why “undetectable” is a claim no steganography tool should make.
Think Twice Before Claiming AI Work as Your Own — What C2PA Content Credentials Prove
AI tools increasingly sign what they create, and editing the file rarely erases it. How C2PA content credentials work, who’s using them, and where LinkedIn fits in.
A Complete Guide to JPEG Forensic Analysis
Every JPEG carries six forensic signals beyond what you see in a photo viewer. This guide covers all of them and explains what they reveal in combination.
How to Forensically Detect AI-Generated Images — No Detection Model Required
Three methods now exist for identifying AI-generated images. Two of them fail when the creator says nothing. This article covers all three — and goes deep on the forensic signals that work regardless of what any label says.
Can you get metadata from a photo someone sent you?
Whether you can read metadata from a received photo depends on the channel, not the tool. Email and AirDrop pass the file unchanged. Messaging apps re-encode — EXIF is gone, but a compression signature takes its place.
The EXIF thumbnail that gave the edit away
JPEG, WebP, PNG, HEIC, AVIF, and camera RAW files embed a firmware-generated thumbnail before any editing software has touched them. When software re-saves the image without updating the thumbnail, the mismatch is forensic evidence — one of 60+ forensic checks snapWONDERS runs automatically on every photo and video.
MakerNote forensics: what camera manufacturers hide inside every photo
MakerNote is the manufacturer-written EXIF block that standard stripping tools typically leave intact — carrying lens serial numbers, shutter counts, burst UUIDs, and capture context that survives a standard privacy scrub. Its absence from a known device is equally revealing.
How to tell if a photo has been edited
How to tell if a photo has been edited: forensic analysis reads compression signatures, manufacturer metadata baselines, and thumbnail mismatches across JPEG, WebP, PNG, and video.
Your camera’s fingerprint survives EXIF stripping — here’s how
Stripping EXIF metadata doesn’t make a photo anonymous. DQT quantisation tables and Huffman coding tables are baked into the JPEG compression — a permanent camera fingerprint that survives every standard metadata strip. Two of the 60+ forensic checks snapWONDERS reads from every image.
How I built a JPEG encoder fingerprint database from scratch
How I built the DQT encoder fingerprint database inside snapWONDERS — deriving libjpeg tables mathematically from the ISO standard, and running a live accumulation pipeline with a trust model to guard against faked metadata. And what this has to do with Vaultify.
Huey’s vision — and the decade of work to honour it
The story behind Vaultify — a father building a steganography platform in memory of his son Huey, who lost his sight at four and dreamed of photos that carry hidden stories.
