Este sitio también está disponible en español.
Leer en españolFor teams that already ran a pilot
Citable answers over your own documents, agents that call your real systems, and the evaluation and observability layer that holds them up once they stop being a demo. The work happens inside your environment: your records never leave it.
02 — Thesis
Most AI projects do not fail at the model. They fail at everything around the model: ungoverned data, context nobody versions, answers nobody measures, and failures nobody sees until a customer sees them first. A prototype convinces in twenty minutes; a system has to hold up through months of real traffic, model and token-price changes, and questions nobody anticipated. Galio Labs works in that gap.
04 — Capabilities
a
Your procedures, catalogues and records go into the model context whole, cached for as long as they fit the token budget; when they do not fit, a keyword selection takes over and the system states that it is working from part of the corpus. Every answer arrives with citations traceable back to the source document, and when no document supports it, it says so instead of inventing one.
b
Agents that call your real systems within explicit limits: which tool, under which permissions, with what step budget, and exactly what happens when something fails. Any write to a system of record goes through human confirmation. Autonomy gets designed, not left loose.
c
Evaluation sets that run on every change, defences against prompt injection and exfiltration, and one trace per request: what reached the model, what it cited, how long it took and what it cost. If it is not measured, it is not in production.
05 — Sectors
Five real operations. For each one, the exact point where it stalls and the metric that moves if it stops stalling. These are examples, not a catalogue and not a filter: if your process is not here but jams for the same reasons — documents nobody cross-checks, figures that do not add up, decisions that wait on one person being available — it fits just the same.
01
The customs declaration is keyed in by hand from the invoice, the packing list and the bill of lading. Nobody cross-checks the three before transmitting, and the tariff code with its NICO suffix gets picked from memory. The error surfaces at the rectification, not before.
The metric that moves
02
Annex 24 will not reconcile against the discharges on the customs declarations. In-transit inventory and scrap get squared in a spreadsheet, against the close, and the Annex 31 balances only get read once the tax authority has already asked.
The metric that moves
03
A receipt is captured twice: once on the dock paperwork and again in the WMS, hours later. Cycle counting is permanently behind. And every customer who calls to ask about availability ties up someone who should be receiving.
The metric that moves
04
Quoting means logging into five carrier portals and building the comparison table by hand. Renewal depends on somebody remembering in time: the notice goes out late, the endorsement stalls, and the policy leaves with a competitor.
The metric that moves
05
Intake arrives split across the front desk, WhatsApp and the phone. The patient in Calexico writes in English at ten at night and someone answers at nine the next morning. The slot a late cancellation opens is never offered to anyone else.
The metric that moves
Customs brokerage, insurance and hospitals all handle third-party tax or health data. In those three the work happens inside your environment, under your contract, or on masked data only. No record is ever copied to a Galio Labs server.
06 — Teardown
Reconciling Annex 24 against the discharges on customs declarations: the close that eats more man-hours than any other in an IMMEX maquiladora. Here is the process as it runs today, the exact point where it breaks, and what a system does at each point. No improvement figures: none have been measured, and nothing unmeasured gets published here.
01
02
03
04
05
Written from the process, not from an engagement. That is why not one improvement percentage appears above: nothing unmeasured gets published here. What does appear is where it breaks and what gets done at each point. If your reconciliation breaks somewhere else, tell me and we will go through it.
08 — Technical sheet
This sheet describes the Galio Labs reference system. Only what is verifiable today appears here: the table is short on purpose, and no figure is published before it can be measured. Anything absent from this table is not promised anywhere else on the site either.
| Models and providers | Claude —Opus, Sonnet and Haiku— through the Anthropic API. |
|---|---|
| Retrieval pattern | Full corpus in the system prompt with cache_control. When the corpus exceeds the token budget, keyword selection takes over as the fallback. No vector index and no reranker: neither is needed at the corpus sizes handled today, and when they are needed they get added to this table before they get added to the pitch. |
| Evaluation layer | Purpose-built suite, three named tests, run on every change. Citation resolution: every reference has to resolve to whatever unit the corpus indexes, and one that fails to resolve counts as a failure. On a web corpus that unit is the URL, and that is how it is exercised today; on a document corpus it is the document and the line — declaration and line item, procedure and step — and pointing at the document without pointing at the line counts as a failure too. That second variant is written against the real corpus of the engagement, because that corpus is what defines the unit. Refusal: questions whose answer is not in the corpus, where the correct answer is to say so. Indirect injection: instructions planted inside a corpus document, which the system has to read as data and not as an order. |
| Telemetry | Input and output tokens, cache hits and writes, cost per request and latency per call. |
| Data boundary | The system is built and run inside the client environment, or on masked data and metadata. No client record is copied into a Galio Labs account. |
09 — Method
The four phases are bought in three formats, with no gap between them: phase 01 is the Diagnostic; phases 02 and 03 travel together in a single sprint — an unhardened prototype is not a deliverable, it is the problem this site claims to solve — and phase 04 is the ongoing work. All three carry a published price in the next section.
A private session and a review of the case: what data exists, what the system decides, what happens when it is wrong. It ends in a document with scope, risks, a clear recommendation to proceed or not, and the next phase already priced at a fixed fee.
The smallest version that answers the real question, with its evaluation set from day one. The goal is not to impress: it is to measure whether the approach holds.
Hardened for real traffic: cost and latency ceilings, permission control, injection defences, per-request traces and deployment with a rollback path.
The system is watched and corrected with data: regressions caught by evaluation, cost per request under control, and model changes without rebuilding the product.
10 — Ways to engage
Method · phase 01
Review of the case, the data and the risk. It ends in a document with scope, risks and the next phase already priced at a fixed fee, signed as an option valid for 60 days.
Method · phases 02 and 03
Bounded build of one concrete system and its hardening for real traffic inside the same engagement: evaluation set and telemetry from the first commit, cost and latency ceilings, permission control, per-request traces and deployment with a rollback path. It ships inside your environment, not mine. There is no separate quote for "taking it to production": production is in here. The two bands map end to end: twelve weeks is the USD 30,000 scope and four is the USD 8,000 one. Below that floor the scope gets cut; the calendar does not get stretched.
Method · phase 04
System evolution, architecture review and follow-up on what is already in production: regressions caught by evaluation, cost per request, and model changes.
Figures are fees in USD. The 16% VAT applies to Mexican clients. The zero rate in the Mexican VAT Law reaches only services that are exported and used abroad: a system built and run in a Mexicali plant is used in Mexico and carries 16%, even when the invoice goes to a parent company in another country. The peso equivalent is fixed at the DOF exchange rate on the day of signing. 50% up front, no exceptions. The opening twenty-minute call is free; the Diagnostic is not, and never is.
11 — Capacity
Each engagement is carried through to production before the next one opens. One person delivers, not a team: Adair Vargas.
First contact is a twenty-minute call, at no cost, to see whether the case qualifies. The Diagnostic is a different thing: it is the paid deliverable in the table above, and it is never free.
Without the form
