Human Agent
EN
Book a demo
Use cases

Five fields where a wrong figure is expensive.

SynaLemma is built for decisions that depend on a figure: a production plan, a recall, a reimbursement, a regulatory filing. Below are five fields where that is the case, and what SynaLemma changes in each.

01
Industry & supply chain

Inventories, bills of materials, supplier risk.

The questions

  • Which assemblies stop if this supplier fails?
  • What is the real stock of this component across sites?

Why a model alone gets it wrong

A bill of materials is recursive, and plain SQL makes any-depth traversal hard to get right. The ERP underneath is often old, and its column names explain nothing.

What the harness changes

  • Recursion as a base case and a recursive case: a component required at any level, in a few lines.
  • Legacy tables used as they are, with their real names. No migration, no copy.
  • One definition of “available stock”, shared by planning, purchasing and the agents that assist them.

A definition it holds SupplierExposure: the assemblies that stop if a supplier fails, through every level of the bill of materials

02
Aerospace & defence

Traceability and configuration, through every level of the product.

The questions

  • Which serial numbers carry a part from this lot?
  • Which aircraft are affected if this supplier's batch is recalled?

Why a model alone gets it wrong

An as-built structure runs many levels deep, and the answer depends on which configuration was valid on a given date. A model writing joins by hand stops one level too early, and nothing tells you so.

What the harness changes

  • The as-built structure modelled as a knowledge graph over the tables you already have.
  • Recursion through every level, to a depth you declare, so it always terminates.
  • Configurations valid over a period: what was fitted on a given date, not only today.
  • Every answer cites its rule, and the rule has a history in git: what an authority or a customer audit asks for.
  • It runs on your servers, with a model you host. Sensitive data stays where it is.

A definition it holds “Every fitted part traces back to a received lot”, stated once and checked against the data

03
Health & insurance

Reimbursement rules and invoice control.

The questions

  • Is this invoice reimbursable, and under which clause?
  • Which claims this month fall outside what the contract covers?

Why a model alone gets it wrong

The rules are dense, dated and full of exceptions. A model paraphrases them, and its paraphrase can change from one question to the next.

What the harness changes

  • Each clause becomes a named rule, with the period over which it was valid.
  • Assertions state what must never happen, and name the claim that breaks them.
  • Every decision cites its rule: what a patient, a payer or an auditor will ask for.
  • It runs on your servers. Health data stays where it is.

A definition it holds “A claim is reimbursable only if the contract covers it”, stated once and checked against the data

04
Life sciences

Knowledge graphs and analysis over experimental data.

The questions

  • Which compounds act on a target involved in this pathway?
  • Which samples of this cohort meet every inclusion criterion?

Why a model alone gets it wrong

The answer sits several relationships away, across many tables. A model writing joins by hand drops one of them, and the result looks just like a correct one.

What the harness changes

  • Entities and relationships modelled as a knowledge graph over the tables you already have, with no graph database to run.
  • Traversals through as many levels as you declare, so they always terminate.
  • Inclusion criteria written once as named rules, so an analysis can be re-run, reviewed and cited.

A definition it holds “A sample is in the cohort only if its consent is on file”, stated once and checked against the data

05
Industrial finance

Inventory value and production cost, defensible to an auditor.

The questions

  • What was the value of our stock, by site, at the end of the quarter?
  • What does one Cargo C2 cost to build, component by component?

Why a model alone gets it wrong

A component has a standard cost, a last purchase price and a weighted average cost, in three different tables, and only one of them is what finance uses to value stock. The model sees three plausible columns and picks one.

What the harness changes

  • One named rule per figure, that every other rule and every question builds on.
  • Lineage from the answer back to the source tables, rule by rule.
  • Settled definitions are locked, and every change is dated and attributed.
  • Reasoning over time: what was true on a given date, not only today.

A definition it holds StockValueBySite · ∀ s v, StockValueBySite s v → v ≥ 0

The common thread

Your field is not on this page? Look for three signs.

Definitions that matter

A term everyone uses and nobody defines the same way: active, eligible, available, at risk.

Data in relational systems

Tables you already run, often old ones. Nothing has to move for SynaLemma to work on them.

Someone who answers for the number

A person who will be asked where the figure came from, and needs a better answer than “the AI said so”.

Book a demo of SynaLemma
Your case

Show us the question you cannot get a straight answer to.

In a free thirty-minute demo, we look at it with you and tell you whether SynaLemma is the right tool for it.

Book a demo of SynaLemma How the harness works