Skip to content
KaiOne
Cases

Real problems, solved with the numbers in plain sight.

Each case describes the technical problem, how we tackled it and what we measured at the end.

Cases are published identified by industry only. We include no client names, no contract data and nothing that would identify them: confidentiality is part of the service.

01Retail · logistics operationsData platform + applications

Every area with its own truth. One data platform.

The challenge

Shipments, incidents, reassignments and rates lived in separate spreadsheets. No two areas shared a view, and every report arrived with a different number.

How we tackled it

The base first: a canonical model that reconciles the ERP with the sources living outside it, with master catalogs, APIs and role-based security. On top of that, applications by role on the same model. AI went in under shadow mode and only reached production once its suggestions were measured against real operations.

Delivered

  • Canonical data platform integrating the ERP and the external sources
  • Web applications running by role, all on the same data model
  • AI reading evidence documents and suggesting reassignments
  • Move from shadow mode to production with measured accuracy metrics
One model
every area on the same base, each with its own view
Zero
duplicate entries across areas after the unification
Shadow → prod
the AI was measured before it decided on its own

Stack

  • Python
  • PostgreSQL
  • Next.js
  • REST APIs

Evolutionary, stage by stage · in production today

02Public sector · market intelligenceSemantic search + managed operations

We built it, we delivered it — and today we still operate it.

The challenge

Official publications describe the same products with a vocabulary different from the internal catalog. Manual cross-checking let opportunities slip through every day — and once that was solved, the platform running AI workloads needed continuous operations without standing up an internal on-call team.

How we tackled it

Semantic search to read context and synonyms where the text does not match, exact matching where the business admits no ambiguity, and supervised learning that improves with every operator correction. Once the repository was delivered, the same team that built it took over continuous operations: monitoring, verified backups and security posture, reported month to month.

Delivered

  • Hybrid search engine: semantic or exact depending on the field
  • Enterprise repository: code, documentation, pipeline and formal QA
  • Continuous operations with backups verified through restore tests
  • Monthly report on availability, incidents and consumption
98%
exact match where the business allows no ambiguityMeasured performance of the engine we built.
93%
semantic accuracy with synonyms and contextMeasured performance of the engine we built.
1
team: the one that built the engine is the one that operates it

Stack

  • Python
  • Vector search
  • Supervised ML
  • PostgreSQL
  • Continuous SecOps

In production · operated by the same team that built it

03Public sector · document operationsData engineering + AI assistance

From isolated spreadsheets to a governed database that runs with AI.

The challenge

Parallel files, divergent data-entry criteria and no record of who changed what. The assessment ran on the real files and sized the problem before anything was proposed.

How we tackled it

Migration to a governed database with three roles, versioning and an audit log of every change. AI-assisted entry proposes from the source document and the operator approves — under the same controls and permissions as manual entry, so turning it on created no rework.

Delivered

  • Migration of the scattered files into one governed database
  • Governed entry: three roles, versioning and an audit log per change
  • AI-assisted entry with human review and approval
  • AI-first from day zero: same controls to propose, review and apply
Hundreds of thousands
rows entered by hand with divergent criteriaMeasured in the assessment of the real files — the starting pain, not the result.
More than a third
duplicate records: the same file entered several timesMeasured in the assessment of the real files — the starting pain, not the result.
One database
replaces the parallel files each area kept on its own

Stack

  • PostgreSQL
  • Python
  • Next.js
  • LLM with human review

In production with AI-assisted entry live

04Enterprise cloud · production environmentFinOps

From audit to automated FinOps operations in 12 weeks.

The challenge

Spend rising with no explanation by area, incomplete tagging, and reservations that no longer matched what was actually running.

How we tackled it

Three phases: understand, optimize, operate. The first low-risk quick wins shipped in week 1, and the program closed with handover and certification of the client's internal team.

Delivered

  • Maturity assessment and technical audit of the spend
  • Opportunity matrix by impact and effort
  • Automated re-tagging with our own scripts deployed in the account
  • FinOps playbook and 12-month plan, with the internal team certified
12 weeks
from audit to automated operations, in 3 phases
Week 1
first low-risk quick wins applied
Handed over
playbook and certification to the internal team

Stack

  • Azure
  • Python
  • Tagging automation
  • Budgets and alerts

Program closed · operations handed to the internal team

Next step

Bring the problem. We hand back opportunities.

A 60 to 90 minute session with your team, free and without touching your systems: you bring the problem and we come out with the framing and the questions still missing. Then we pick one or two areas and measure them, still with no access to production. Only then is there a proposal with fixed scope and price.

Book a discovery session

60–90 min · Free · No commitment

  1. 01

    Discovery session

    60 to 90 minutes with your team. You bring the problem; we come out with the framing and the questions still open.

  2. 02

    Preliminary diagnostic

    We pick one or two areas of value and measure them. No production access, no cost.

  3. 03

    Formal proposal

    Scope, plan and fixed price. With the math in the open and no fine print.

WhatsApp