Hearth · Human-centered AI
The fire is new. The discipline isn't.
AI is the new fire — enormous energy, dangerous loose, transformative held. Hearth builds the structure that holds it: systems your analysts, operators, and auditors can actually sign off on.
Warmth, engineered.
Every civilization organized itself around the same object: a contained fire. The hearth was never the flame — it was the architecture that made fire safe enough to live with. Around it, people cooked, decided, taught, and stayed warm.
That is our work with AI. We start where the stakes are real — oilfield paperwork, service operations, credit files — and we build the kind of intelligent system a careful company can live with for decades. Not fire for its own sake. The stone that holds it.
One practice, three postures.
Advisory that maps before it builds
We sit with your team, walk the workflow on paper first, and find the one place AI genuinely earns its keep. The AI Modernization Audit is where every engagement begins.
Tools you run yourselves
Mantle, Kindling, Andiron, Dispatch — an ecosystem that turns the practice into something your own people operate, audit, and extend.
Human-in-the-loop as architecture
Sign-off is not a checkbox we added; it is the load-bearing wall. Your analysts stay the authors of every decision. The system gives them leverage and keeps the receipts.
Three ideas govern every decision.
Warmth is engineered.
A hearth is masonry before it is comfort. Precision produces warmth, not the other way around. Everything we ship is geometric first, warm second — never decorative, never cozy for its own sake.
The ember, not the blaze.
A tended ember is the most efficient state of fire: quiet, dense, ready. We look for the smallest automation that changes everything, and we resist the demo that impresses but doesn't hold.
Structure protects.
The arch is humanity's oldest trust technology: it holds weight by geometry alone. Our mark is an arch holding an ember — capability held safely by structure. Human-in-the-loop, drawn as a shape.
Parts of the hearth.
Every product is named for something a hearth is made of — plain, physical, one word. Together they are the modernization stack for companies that measure twice.
Mantle
Platform · consoleThe working surface. Workflows, agents, review queues, and audit trails in one console your team opens every morning — the shelf the whole operation rests on.
See it thinkKindling
Audit · onboardingThe starter kit. Workflow surveys, document samples, and readiness scoring that turn "we should look at AI" into a mapped, priced, defensible first project.
Start the auditAndiron
Integrations · infrastructureThe iron underneath. Connectors into the ledgers, ERPs, and shared drives you already trust — so the fire sits on structure, not on top of your data by force.
Read the docsDispatch
Newsletter · journalField notes from the work. What held, what didn't, and what we'd do differently — written like a letter, sent when there's something worth saying.
Read DispatchWhat we build, concretely.
Workflow survey
We map how the work actually moves — tickets, memos, approvals — before a line of code. The map is the deliverable.
Document intelligence
Field tickets, invoices, covenants, credit files — read into structured lines with the source page always one click away.
Reconciliation agents
Line-by-line matching against rate tables, AFEs, and ledgers. Exact matches post to a queue; exceptions wait for a person.
Review queues
Nothing posts without a named human sign-off. The queue shows what the system did, why, and what it wants permission for.
Integrations
Andiron connectors reach the systems of record you already run — QuickBooks to SAP, shared drives to core banking.
Measurement
We measure analyst leverage — hours returned to judgment — and report it plainly. Numbers get context or don't appear.
Watch the work, not a demo reel.
A morning in Mantle: an analyst hands the system a reconciliation, the system shows its reasoning, and everything ambiguous waits for a person. This is the whole philosophy, running.
Filled dots are where AI acts. The outlined dot is the load-bearing wall: a person, named, signing off. Every step writes an audit line.
Boring where it should be boring.
The interesting part of our stack is how uninteresting most of it is. Retrieval grounded in your own documents. Models treated as replaceable parts, never as the foundation. Audit lines the way a ledger writes them: every action, every source, every signature.
We are model-agnostic on purpose. The best model this year won't be the best model next year, and your workflows shouldn't care. The structure is ours; the fire is whoever burns brightest.
The shape of the work.
Composite scenarios drawn from the patterns we build for. Presented honestly: illustrative, not client results.
From field ticket to finance, without the Friday backlog.
- Crews' paper tickets stop dying in a scanner folder — intake reads them the day they arrive
- Reconciliation runs nightly against rate tables and AFEs; exceptions queue for the analyst
- The analyst's Friday changes from retyping everything to reviewing what actually needs judgment
Credit memos that cite their sources.
- Memo drafts assemble from the ledger, covenant docs, and prior files — every figure links to its page
- The analyst edits judgment, not formatting; the system keeps both versions and who changed what
- Committee reviews arguments with receipts attached, and sign-off is recorded like any other entry
The console has an API.
The API has a conscience.
Everything Mantle does in the console, Andiron exposes as an API — with the same rule baked into the schema: work that needs a human says so, and won't complete without one.
$ curl https://api.hearth.example/v1/runs \
-H "Authorization: Bearer $HEARTH_KEY" \
-d '{ "workflow": "reconcile-field-tickets",
"period": "2026-05", "site": "pad-7" }'
// response
{
"run_id": "run_8Kd2…",
"matched": 197,
"drafted_adjustments": 12,
"held_for_review": 5,
"review_required": true,
"posted": 0,
"audit_ref": "ledger://runs/8Kd2/audit"
}Example only — illustrative API surface, not a live endpoint.
Field notes, sent occasionally.
The ember, not the blaze
Why the smallest useful automation beats the most impressive one, every time it matters.
→ 06 · 2026Ledgers are ground truth
What oilfield paperwork taught us about retrieval, and why the source page is the product.
→ 05 · 2026The review queue is the product
Designing for sign-off: what changes when you treat human judgment as the feature, not the fallback.
→Working notes.
Small, honest studies from live systems. Shared when the evidence is worth more than the polish.
Measuring analyst leverage
A practical definition of "hours returned to judgment," and why we refuse vanity metrics.
Failure modes of document extraction under rate-table drift
What breaks when the vendor updates the rates and nobody tells the model.
Quiet interfaces
Attention budgets in back-office AI: when the right amount of interface is almost none.
Sit closer to the fire.
A small team doing patient work in impatient territory. We hire people who like real workflows, real constraints, and shipping things a CFO can explain.
Bring us the workflow that keeps you up at night.
The first conversation is a working session, not a pitch. Tell us which workflow hurts, and we'll tell you honestly whether AI belongs in it — sometimes the answer is "not yet," and we'll say so.
We reply within two working days. No newsletters unless you ask for Dispatch by name.
Received.
Thank you — we'll read it carefully and reply within two working days. The fire's on.