AI engineering
Copilots, document intelligence and structured extraction, built for the inputs you actually get rather than the ones in the demo. Most generative features fail on the messy 20% of traffic, so that is where we start.
We ask for the inputs you are embarrassed by. The messy fifth of your traffic is where generative features fail, and it decides whether this is a prompt, a pipeline or a trained model.
The context and the output schema are designed before the prompt — what is retrieved, what is summarised, what is left out. The schema is what turns a paragraph into something your code can rely on.
Prompts and context live in version control with the rest of the system. Changing either is a pull request that re-runs the evals, not an edit in a console nobody can diff.
A scored set built from your real documents gates every release. We report the failure classes alongside the score, because the shape of what it gets wrong is what tells you whether to ship.
Low-confidence output goes to a person by design, and that queue is the next training set. The feature improves because it is being used, not because somebody tuned a prompt.
Extraction against a schema, with anything under the confidence threshold routed to a review queue instead of into your database.
"""Extraction against a schema, with a review path when unsure."""
from anthropic import Anthropic
from pydantic import BaseModel, Field
client = Anthropic()
class Invoice(BaseModel):
supplier: str
total_cents: int = Field(ge=0)
currency: str = Field(min_length=3, max_length=3)
confidence: float = Field(ge=0, le=1)
response = client.messages.create(
model="claude-sonnet-5",
max_tokens=1024,
tools=[
{
"name": "record_invoice",
"description": "Record one invoice exactly as written.",
"input_schema": Invoice.model_json_schema(),
}
],
tool_choice={"type": "tool", "name": "record_invoice"},
messages=[{"role": "user", "content": document_text}],
)
invoice = Invoice.model_validate(response.content[-1].input)
# The messy 20% goes to a person, not to a guess.
if invoice.confidence < 0.85:
review_queue.add(invoice, source=document_id)A retrieval assistant over clinical policy and procedure documents, answering staff questions with span-level citations and refusing to answer when the documents do not support one.
A document-extraction pipeline for a retail group, classifying and extracting structured records from supplier invoices and delivery notes arriving in every format a supplier felt like using.
Usually not first. Better context, a tighter schema and a stronger retrieval step beat fine-tuning on most tasks and cost far less to maintain. We fine-tune when the evals show a specific gap that prompting cannot close — typically format adherence or a narrow domain vocabulary.
Generation is constrained to a schema and validated before it leaves the service, with an explicit failure path when the model cannot comply. Downstream code receives a typed record or a handled error, never free text it has to parse.
We set a cost-per-run target during Discover and hold the build to it — through model choice, context trimming, caching and routing cheap requests to smaller models. You get the projected unit economics before the build rather than a surprise on the first invoice.
Yes. Open-weight models self-hosted with vLLM, or a provider inside your own cloud account such as Bedrock or Vertex, when data residency or procurement rules that out. We size the trade-off in latency and cost before you commit.
A first call is 30 minutes and costs nothing. Bring the problem rather than a spec — the useful part is usually working out whether this is the right shape of solution at all.
contact@algologix.coWe reply within 24 hours.