A regulatory summary. A press release buried in the wrong tab. A deal memo nobody wants to touch. Elena gets through all of it before 3:11pm — without filing a single IT ticket.
A story about finding something that shouldn't be there.
An artifact from the wrong century. No internet. A funding committee call at seven. Dr. Reyes has one night, forty gigabytes of site data, and a hunch she needs to prove before morning light.
Document RAGOffline — no internetImage analysisLocal LLMs
Four offices. Zero IT tickets. Kai configures once, exports a JSON file, and emails the whole AI stack to a new team three time zones away. They're running client deliverables before lunch.
The email from her manager came in at 8:47 — a regulatory summary. Needs to be on the VP’s desk before noon. She opens the tools she always opens — three tabs, the usual stack — and feels the familiar weight of it. Cut. Paste. Reformat. Hope the context survives the next session.
It doesn’t.
She closes everything and opens Nemilia.
She picks the research workflow from the drop-down. Types one sentence. Clicks run.
The screen comes alive. A web search agent fires first, pulling sources she would have spent twenty minutes finding. Then three agents at once, each running its own delegated task, each working a different angle of the same question. It feels less like a tool and more like finally getting the experts the boss keeps saying they can’t afford.
A review pass. A synthesis. Then a clean document sitting in her results, ready to export.
She looks at the clock. 9:19.
She exports to Word and sends it before her coffee goes cold.
Then she sees it.
A press release. A company she’s been watching for six months — a potential counter-party in a deal her team has been quietly circling. The details are buried in the third paragraph the way important things usually are. She right-clicks on the page.
Send to Nemilia.
The extension takes it. The whole page — text, structure, context — lands in her Captures inbox without her changing tabs. She right-clicks a chart further down. The image follows. She finds a related filing two clicks away and sends that too. Captures, tagged, queued, waiting.
She goes back to Nemilia. Clicks on Captures. Three items sitting there like evidence on a desk. She looks them over, routes the press release and the filing to her research workflow with a single click, sends the image into the library where it gets analyzed and indexed automatically.
Then she clicks run and goes to get more coffee.
By the time she’s back, it’s done.
Then the knock on her door.
Her colleague sets a printed memo on her desk. Forty pages, tight margins, the kind of thing that announces itself. Deal memo. He doesn’t need to say the rest. She already knows. Client sensitive. Conflict protocols. The IT policy she has memorized the way you memorize fire exit locations — never expecting to need it, then suddenly needing it all at once.
She used to have one move here. Read it herself. All forty pages, a highlighter, a notepad, the particular exhaustion of trying to hold a deal structure in your head long enough to write something coherent about it. Or she’d file a request and wait for a tool that fit 70% of what she actually needed and required three follow-up emails to get.
She opens the agent builder.
Two agents. She names them. Writes their instructions the way you’d brief a smart colleague on their first day — what to look for, what matters, what to flag without being asked. Low temperature. She’s not looking for creativity right now.
She drags them into a workflow. Adds a checkpoint on the regulatory agent — she wants to see that output herself before it goes anywhere.
Then she opens the provider selector and switches to her firm’s internal LLM. Nemilia finds it on the network without being asked. She selects it. The workflow doesn’t change. The pipeline doesn’t change. The forty pages load into the library and she clicks run, and she watches her own custom AI team go to work on a problem she built the team for twenty minutes ago.
Then she gets to page 31.
The counter-party. She sees the name and sets down her pen.
This one doesn’t go to the internal server either. Conflict protocol. She knows the rule. She’s seen what happens when someone forgets it.
She opens the provider panel. Switches to the in-browser option. Loads an LLM her security team approved last quarter. It downloads into the browser tab, caches, and then it’s just there — running on her machine, on her GPU, making no calls to anything anywhere.
She puts the relevant section in. Asks her questions.
The answers come back. They’re hers. They live nowhere else.
Her laptop makes a short tone — the pipeline pausing itself. The regulatory risk output is on screen, waiting. She reads it slowly. Checks the citations against the source documents in her library. They hold up. She clicks Approve.
The pipeline moves on.
The synthesis agent finishes. She exports to PDF.
It’s 3:11.
She sends the report and sits for a moment in the quiet of a problem that got solved without asking anyone’s permission, without a policy violation, without leaving a trail somewhere she’d have to explain later.
Tomorrow she’ll open Nemilia and her agents will still be there. The workflow she built today will be waiting. Her document library, intact. The images she captured this morning, indexed and searchable, part of the record.
She closes her laptop.
She used to think AI was only for the guy on her team who sent Slack messages about it at 11pm. Now it just feels like part of the job.
The artifact came out of the wrong layer.
Dr. Reyes knows what it is. She’s held enough of them to know. The glaze, the profile, the way the handle seats — twelfth century. Maybe eleventh. The stratigraphy says fourteenth. The stratigraphy has been consistent for three seasons. The stratigraphy does not make mistakes.
She sets it down on the examination table and looks at it for a long time.
The funding committee call is at seven. It is currently ten-fifteen. The site is three hours from Amman. The satellite connection dropped four days ago and nobody is coming to fix it before Friday.
She opens Nemilia.
She built the library in the two weeks before departure. Methodical. Everything she thought she might need and a few things she hoped she wouldn’t. Three seasons of excavation logs. Soil analysis reports. Carbon dating results. Site photographs — every one dragged into the library and vision-analyzed before she left, so they’d be searchable by what was in them, not just what they were named. Published literature: 340 papers, PDFs imported one folder at a time, chunked and embedded against the day the internet became theoretical.
She had done this before. The internet is always theoretical past the second checkpoint.
She opens the agent builder. Three agents. She writes each one herself — a literature agent told to pull known parallels from the academic papers, a site agent pointed at her excavation logs and dating reports, a synthesis agent told to reconcile what the other two find and list explanations in order of evidential support. She drags them into a sequence, sets temperatures low, and types one question into the workflow prompt: What are the known explanations for chronological displacement of twelfth-century Levantine ceramics in fourteenth-century strata?
She makes tea on the camp stove and sits in the doorway while they work.
The output lands seventeen minutes later. Six possible explanations ranked by how well her own documents support them. She reads slowly, the way you read something when you’re hoping for a specific answer and trying not to want it too badly.
Explanation four stops her.
Intrusive burial. A later interment cut into an earlier layer, disturbing the surrounding matrix. It happens. It leaves traces. She knows exactly which photographs would show it if it were true.
She builds one more agent. Instruction specific enough to brief a first-year doctoral student: Review all site photographs from Grid C, Trench 7, Layers 4 through 6. The images are in the library. Identify any soil discontinuities, cut features, or color changes inconsistent with natural deposition. List findings by photograph name and describe exactly what you see. Do not interpret — describe.
Temperature: zero.
Then she goes outside and looks at the stars for a while, because there is nothing else to do and because on nights like this, in places like this, it feels wrong not to.
Her laptop makes a short tone.
She reads the output standing up, still holding her tea.
Three photographs flagged. Soil discontinuity in the northeast quadrant of Layer 5. A slight color shift in the matrix. A line that doesn’t follow the surrounding deposit — the kind of thing you miss when you’re moving fast, when the season is ending, when the light is bad.
She pulls the photographs herself. Opens them in the library preview. Zooms in.
The line is there.
She spends the next two hours building the case. Cross-referencing the findings against the site logs. Checking the dates of those specific photographs — who was on site, what conditions, what they noted or didn’t note. She builds a second workflow: two agents in sequence, the first mapping the spatial relationship between the cut feature and the artifact’s find-spot, the second checking every prior season’s notes for any mention of disturbance in that grid. The documents are all in the library. The agents work through them while she annotates photographs by hand the way she was trained to, back when this was the only way.
At 2am she exports to PDF. Fourteen pages. Sources cited. Photographs annotated. Confidence: moderate, pending physical re-examination of the cut feature in morning light.
She sets an alarm for five-thirty.
The funding committee call goes long. They ask good questions. She has answers for all of them.
After the call she walks back to Trench 7 with a trowel and a fresh eye.
The line is there. Right where the photographs said it would be.
She kneels in the dirt and starts to work carefully, the way you work when you think you’re about to find out what happened eight hundred years ago in a place no one has looked since.
She used to think the hard part was the digging.
Kai doesn’t think about software deployment anymore.
He used to. He has a list in his head — not written down, because he never needed to write it down, because it cost him enough that he memorized it the way you memorize things that hurt. Forty-two days to procure a software contract for his second office. A contractor at $150 an hour to configure a tool that should have taken twenty minutes. Two good people who left because they needed to move faster than the infrastructure would let them. A three-day thread with IT about a Python environment that kept breaking on one specific machine, on one specific version of Windows, for reasons nobody ever fully explained.
He doesn’t think about any of that anymore.
The third office opens on a Tuesday.
The Friday before, Kai sits down for ninety minutes. He’s been building the workspace for a week. Not because it’s hard — because he’s particular. Each agent gets a clear brief: what to look for, what to flag, how to hand off to the next one. A research agent. A document review agent. A competitive intelligence agent. A synthesis agent that outputs in the format his reports already use.
Each one a JSON file. Each one done in a sitting.
He exports the workspace. One JSON file — agents, workflows, provider configuration, MCP server connections, all of it. He attaches it to an email along with the HTML file, which is the entire application, the size of a floppy disk, not that anyone on his team has ever held one, a fact he has made peace with.
He types four words in the body: Open both. Import workspace.
He hits send and closes his laptop.
Monday morning, 9:11am. The new office lead — three time zones away, laptop just out of the box — opens Chrome. Opens the HTML file. Imports the workspace JSON. Every agent Kai built appears in her interface, named, configured, ready. The MCP connections resolve automatically. She enters her API key once, thirty seconds, done.
She runs her first workflow at 9:26.
No ticket. No installation. No IT thread. No contractor. No Python.
Kai gets a Slack message at 9:38: this is wild.
He knows.
He remembers the day he stopped thinking of AI tooling as something his team had to be onboarded to and started thinking of it as something he configured once and sent. The workflow JSON is in a shared folder now. When he refines an agent — tightens the instructions, adjusts the temperature, adds a step — he exports and shares the update. Everyone gets it. It takes him four minutes.
The whole stack fits in an email. It runs on any machine with a browser. If someone gets a new laptop, they’re back up in the time it takes to open two files.
He shipped a GPU laptop to his Singapore hire last month with the workspace JSON already imported, a WebLLM model pulled from HuggingFace and cached locally, API key already entered, every workflow already tested. She was running client deliverables on her first day. Not her second day. Not after IT cleared the environment. Her first day, before she’d finished unpacking.
He has four offices now.
The bottleneck used to be setup. Now the bottleneck is hiring, which is a better problem.
His next office opens in six weeks. He already has the workspace ready.
