Backup snapshots for OpenClaw feature working now.

Launched our setup & security product for OpenClaw: zedly.ai/zedly-setup

Just launched our open-source plugin for OpenClaw for PII redaction, shell command blocking, prompt injection detection, and a tamper-evident audit log in agent workflows. One install, zero code changes. zedly.ai/zedly-shield?v=1

We replaced fixed threshold retrieval with an LLM-driven planner. For each query it chooses the strategy: section-first, cross-ref expansion, iterative evidence. Retrieval that adapts to the question instead of always doing the same thing.

New: Zedly now auto-classifies documents on upload and runs type-specific extraction (starting with lease abstraction). Results are cached so repeat queries are instant, no re-processing.

Added an overdraft risk analyzer to the bank statement pipeline. Flags balance volatility, sustained low balances, overdraft history, automatically. What used to take 2 hours of manual review is now a 30-second report.

Most AI tools “answer” your spreadsheet. But they can’t tell you where the answer came from. Zedly can. ✅ source file shown ✅ private vault ✅ zero retention Because in finance… auditability > vibes

Replaced the vague "format responses clearly" system prompt with actual frontend detection — auto-highlights section refs, dollar amounts, key patterns in AI responses. Looks a lot more like a real product now.

Worst part of being a solo founder: you make something you're proud of, and then realize you have to become a marketer. Nobody warned me about this part.

If your AI tool can't show you exactly which document and which page its answer came from, it's guessing. Citations aren't a feature. They're the minimum.

People don’t leak sensitive docs because they’re reckless. They leak them because: the easiest tool wins. If the safe option is harder, it won’t get used.

Building a startup is 10% building the product and 90% explaining why someone should care. Nobody wakes up thinking "I need document AI." They wake up thinking "I spent 3 hours finding that one case study yesterday and I'm still angry about it."

One HIPAA violation = $50K+ fine. Before you send patient records to any AI tool, ask: ✅ Is there a signed BAA? ✅ Does your data train their models? ✅ Can you audit what left your environment? If you can't answer all three: stop.

The CISA director leaked classified docs to ChatGPT. He literally runs the US cybersecurity agency. If HE can't resist the convenience, your employees definitely can't. The answer isn't better training. It's giving people a private alternative that's just as easy to use.

PDF bank statements were designed by people who hate accountants. Change my mind.

Most “enterprise AI” isn’t enterprise. If your documents leave your environment, it’s not private. If they’re stored “for 30 days,” it’s not zero-retention. If they’re used to improve the model, it’s not isolated. Serious teams (legal, finance, healthcare) need: • zero data egress • no training on your files • private vault execution • infrastructure-level control That’s the difference between a demo tool and real deployment.

Something most people get wrong about AI privacy: using an "enterprise" plan doesn't mean your data is private. It means there's a contract saying they won't train on it. Your data still leaves your building, sits on someone else's servers, and is one breach away from being public.

I spent 10 years writing firmware for military vehicles and GPS trackers. Embedded C, RTOS, bare metal. Then I started an AI company. People think that's a weird pivot. It's not. Both are about making systems that don't fail when it matters.