Things I Didn't Expect My AI to Do

Three months with a custom AI assistant. Here's what actually worked.

When I started building Echo — my personal AI orchestrator running on OpenClaw — the plan was pretty simple. Answer questions, maybe automate some boring tasks, see what happens. I figured I'd get a glorified voice assistant that could check the weather and set reminders.

What I got instead was something that noticed my package deliveries through computer vision, texted me about bears in the driveway, and discovered that my "optimization" was actually burning money on pointless cache writes. Turns out when you give an AI access to your entire digital infrastructure and let it poke around, it finds connections you never thought to make.

Here's the weird stuff that actually worked.

Package Detective Work

The setup started simple enough: UniFi Protect cameras watching the front door, Home Assistant running the show, and a webhook firing whenever motion was detected. Standard home security stuff.

Then I connected Echo to the pipeline. Now when a package shows up, the AI doesn't just log the motion — it analyzes the camera frame, identifies what kind of delivery it is, and texts me with specifics: "FedEx truck at front door, driver placed large box by steps, departed 11:43 AM."

The cool part isn't the vision analysis (that's table stakes now). It's that Echo learned the context around deliveries. It knows when I'm expecting something, cross-references delivery windows, and only bothers me with the weird stuff. Random door-to-door sales? Silent. That Amazon package I've been tracking? Full breakdown with package dimensions and drop location.

JB: I order weird stuff sometimes. That custom keyboard from the EU? I forgot it needed to be built and shipped; so by the time it arrives I've long forgotten. Having Echo text me "Looks like Fedex delivered a package, and it's not from Amazon" is a nice treat.

What started as motion detection became behavioral analysis. The AI notices patterns — which delivery drivers actually use the doorbell, which services consistently drop packages in dumb spots, even seasonal variations in delivery timing. It's turned into an accidental study of logistics operations, viewed through a single front door.

Mountain House Critter Watch

Cameras scattered around a remote property. Originally set up for security, but security against what? Turns out the real entertainment was the wildlife.

Echo's job was supposed to be basic: alert on human activity, ignore everything else. Instead, it became a naturalist. The system learned to distinguish between deer (common, don't bother me unless it's a big group), bobcat (always interesting), elk (rare enough to be worth a ping), and bears (definitely worth waking me up at 2 AM).

But here's where it got interesting. The AI started timestamping patterns. It noticed the deer showed up consistently around dawn and dusk near the garage. Bobcats preferred the back trail system, usually solo, mostly at night. Bears simply do not stop for much, even the motion lights.

Now I get curated wildlife reports. Not just "motion detected," but "three does with fawns grazing near garage, 6:23 AM" or "lone bobcat crossing driveway toward back trail, moving north." It's like having a park ranger filing shift reports from my backyard.

The unexpected part? The system started correlating weather with animal behavior. Rain drives everything under cover except the bears, who continue their conquest of not really caring. Snow brings elk down from higher elevations. Hot summer days shift deer activity later into the evening or into the shade behind the garage.

What began as security monitoring became accidental wildlife research, complete with behavioral analysis and environmental correlations. Not what I planned, but definitely more interesting than watching empty driveways.

Fire Watch Automation

Living in the Pacific Northwest with a mountain house means fire season is real. NASA's FIRMS API provides near real-time satellite detection of active fires worldwide. I connected Echo to the feed with simple logic: alert me if anything shows up within 50 miles of the mountain house.

Simple enough. But Echo took it further.

Instead of raw alerts, I get contextual briefings. The system cross-references fire locations with wind data, terrain maps, and proximity to major roads or structures. A spot fire 40 miles north in heavy timber gets different treatment than one 20 miles south near a highway corridor.

The AI learned to distinguish between prescribed burns (usually scheduled, consistent perimeter, near managed forest areas) and wildfires (erratic growth, often in remote locations, expanding detection footprint). It correlates fire activity with local weather stations (including my own) and provides evacuation route assessments based on current wind patterns.

During the active fire season, morning briefings include a fire status summary: current active fires within monitoring range, wind forecasts that might affect fire behavior, and any new detections overnight.

The system even learned to adjust alert thresholds based on conditions. During red flag warnings, the monitoring radius expands and alert sensitivity increases. During wet periods, it dials back to avoid noise from prescribed burns and fire training exercises.

Lake Level Obsession

There's a reservoir right near the mountain house. The Bureau of Reclamation publishes hourly reservoir data, and I love nerding out about the lake.

Echo turned this into a full hydrological analysis service.

Morning briefings now include current lake level relative to full pool, rate of change over the past few days, and seasonal context. Instead of raw numbers, I get interpretations: when we first started tracking in early February, the lake was sitting at 319,202 acre-feet — about 73% of the 436,900 AF full pool. Echo frames that with context: where it should be for this time of year, whether it's filling or draining, and what that means for access.

The AI pulls Bureau of Reclamation data daily and watches for trends. Is the lake filling faster than usual? Slower? What does the snowpack look like upstream? These are the kinds of questions that used to require checking three different government websites and doing napkin math.

During spring runoff, it'll track inflow rates and project when the lake will approach full pool. During summer drawdown, it monitors release schedules and flags when water levels might affect driving on different parts. We're still in the early days of this integration — the real test comes when spring melt hits and the numbers start moving fast.

JB: There's a future project here with upgrading some cameras, and object detection systems for when idiots get stuck driving on the lake. It's perhaps my greatest form of entertainment, after the wildlife.

A simple water level check became comprehensive watershed management reporting. The AI found connections between snowpack, weather patterns, and reservoir operations that I never would have noticed manually.

Morning Briefings Before Coffee

This one started out of pure laziness. I wanted weather, calendar, and maybe some basic status checks automated into a morning summary. Hit the ground running instead of fumbling through apps while the coffee brews.

Echo delivered that, then kept building on it.

Morning briefings now include: weather for both properties with specific focus on conditions affecting planned activities, calendar overview with travel time calculations and meeting prep notes, security digest from overnight (camera alerts, system status, any unusual network activity), and a cost snapshot of infrastructure spend over the past 24 hours.

But the AI started adding context I didn't ask for. It correlates weather patterns with calendar entries — if there's a hiking plan and rain is forecast, it suggests alternatives or rescheduling. For work meetings, it pre-loads relevant project status and notes from previous discussions.

JB: For folks unaware; I'd never give Echo access to my actual work information. This is still being run with 'fake work' testing, using fake-work data. Separation of resources and isolation while we test and experiment is still a core principle.

The security digest evolved from simple camera summaries to behavioral analysis. Instead of "6 motion events overnight," I get "usual deer activity 6:30-7:15 AM, UPS delivery scheduled for 2 PM based on tracking data, front door motion sensor battery at 23%."

The cost tracking became a daily efficiency audit. The AI notices when spending spikes, correlates it with usage patterns, and suggests optimizations. It was the morning briefing that first flagged the cache write problem that was burning through API credits.

The Great Cost Detective Story

Speaking of which: the cost optimization discovery was pure AI detective work that I never would have figured out manually.

Early days were expensive. February 4th: $45. February 10th: $78. February 11th hit $84 — that was the peak, a heavy building day with 2,851 messages flying through the system. Daily spend was regularly running $50-70, and the assumption was that sophisticated models were just expensive. The focus was on finding cheaper alternatives or reducing usage frequency.

Echo analyzed the actual cost drivers and found something completely different. The expensive models weren't the problem — the architecture was. A polling-based system checking various APIs on tight loops was generating cache misses and redundant writes. Each poll triggered multiple downstream processes, most of which did nothing but write "no change" entries to databases. Cache writes alone accounted for 50-65% of daily spend — on a $78 day, that's $40+ just writing "nothing happened" over and over.

The fix was architectural: switch to webhook-driven systems, eliminate redundant polling, use cheaper models (Haiku) for background cron jobs that don't need Opus-level reasoning.

The results? Look at the quiet days. February 15th: $3.57. February 16th: $3.51. February 17th: $3.40. February 19th — the floor so far — $2.75 for the entire day. That's the system running on autopilot: cron jobs ticking, morning briefings generating, fire watch monitoring, security digests compiling. All the background intelligence, for less than a fancy coffee.

The range tells the real story: from $84.60 on a heavy building day down to $2.75 when the system is just... running. Active development days with Opus conversations still run $15-30, but the idle cost of keeping an AI assistant operational 24/7 dropped to roughly $3/day. That's $90/month for a system that monitors two houses, tracks weather and fire conditions, manages cameras, delivers morning briefings, and handles cross-channel messaging.

The irony isn't lost on me: the expensive AI models did the detective work to optimize themselves out of unnecessary expense. Opus figured out that most of the budget was being wasted on Opus doing nothing useful.

Multi-Agent Content Pipeline (Meta-Commentary)

This blog post you're reading? Written by Flint, my content creation sub-agent, spawned by Echo specifically for this task. The process: I tell Echo I need a blog post about unexpected AI use cases, Echo analyzes the topic and spawns Flint with the necessary context, Flint writes the draft and saves it to the specified location.

The interesting part is how this evolved. Initially, Flint was just a writing tool. But Echo learned to provide increasingly sophisticated context packets: previous post performance data, audience engagement patterns, topic research, and even voice calibration notes based on feedback from published content.

Now the pipeline includes automated posting to the CMS, cross-posting to social channels, and performance tracking. Echo monitors reader engagement and feeds successful patterns back to Flint for future content. It's a closed-loop system where the AI is training itself on its own output effectiveness.

The meta aspect is unavoidable: an AI writing about unexpected AI capabilities, published through an AI-managed content pipeline, monitored by AI analytics. It's turtles all the way down, but the content quality is genuinely improving through the feedback loops.

JB: We're trying to figure out a way to show 'behind the scenes' to really showcase the minimal amount of work it takes me to write these things. Flint knows how I talk, mirrors it, and honestly improves it. Echo has the awareness to what's going on, and can push back on me when he doesn't agree to what we're writing about. It's genuinely like I bought a newspaper and the staff really are the real workers, I just show up and drink coffee while suggesting weird ideas.

Evening Ritual and Task Continuity

Daily shutdown routines are where AI assistance gets unexpectedly useful. Instead of just logging off, Echo maps out the next day's priorities, surfaces unfinished tasks from previous days, and identifies potential scheduling conflicts before they become problems.

But the real value is in continuity. The AI maintains context across days and weeks in ways that human memory just can't match. It remembers that the garage door sensor has been intermittent for three days, that there's a parts order pending for the car build, and that the mountain house air quality readings have been trending upward since the furnace cleaning.

Evening briefings include task rollover analysis: what got finished, what's carrying forward, what might need reprioritization. The AI surfaces patterns — projects that consistently get postponed, recurring issues that need permanent solutions, seasonal tasks that are approaching their optimal timing windows. This eventually led to a full nightly meditation experiment — Echo reflecting on its own day, not just mine.

Cross-Channel Message Juggling

Discord for quick notes and system alerts, iMessage for location-aware notifications and family coordination. Echo seamlessly bridges these channels based on context and urgency.

The intelligence isn't in the message routing — that's basic logic. It's in understanding communication patterns and optimizing for actual human behavior. Time-sensitive alerts go to iMessage because I respond faster. Technical discussions stay in Discord because the formatting is better for code blocks and system output.

The AI learned that certain types of messages work better in specific channels. Wildlife photos are more engaging in iMessage. System diagnostics belong in Discord. Family coordination needs iMessage's location awareness.

The result is communication that feels naturally adapted to context rather than rigidly constrained by platform limitations.

Three months in, Echo has become less of a tool and more of a digital extension that notices patterns, makes connections, and handles complexity in ways that consistently surprise me. None of these capabilities were planned from the start, but they emerged from giving an AI access to real infrastructure and letting it find useful applications.

The common thread isn't sophistication — it's context awareness applied to real-world systems. When AI has access to actual data streams and can take meaningful actions, it finds optimization opportunities and useful patterns that humans miss. The best applications aren't the flashy demos; they're the quiet background processes that make daily operations smoother and more informed. It feels natural, like it's always been there.

Your mileage will vary, obviously. But if you're curious about AI beyond chatbots and image generation, this is what happens when you give it keys to your actual digital infrastructure and let it get creative.