Cosine and Euclidean stop separating the right records
Catalogs, multilingual corpora, and wide embeddings can look mathematically close while remaining operationally unrelated.
Clustering consultancy for businesses that hit the limits of default workflows
When the data stops behaving nicely, the clustering strategy has to change.
De Novo Clustering is a team-led lab focused on clustering problems that look simple at first and turn difficult in production: non-spherical structure, uneven density, drift, mixed data, sparse text, and distance metrics that stop separating the right neighbors.
The service starts with the brief, assesses the actual geometry and business risk, and returns a plan of action. Clients either receive clustered outputs and reporting, or an approved API that applies the verified methodology to future data.
specialized clustering work
20 years
Direct judgment on each engagement, not anonymous delivery.
how the work starts
Manual first
Geometry, density, drift, and outliers are assessed before a method is approved.
what gets delivered
Validated workflow
Receive clustered outputs and reporting, or a verified API for repeat data.
Default metrics fail quietly when the geometry is wrong. That is usually where specialist review becomes worth the cost.
Clusters are judged in business terms: whether they change a decision, survive challenge, and remain useful after deployment.
The workflow is validated before it is scaled. Clients do not inherit a black-box clustering guess.
Why businesses hand this over
Why businesses hand this over
Most teams do not arrive saying “we have a clustering problem.” They arrive saying the groups are unstable, the neighbors feel wrong, the centroids are swallowing outliers, or the old workflow falls apart as soon as the data shape changes. This section is written to sound more like those conversations.
What can stay in-house and what is worth delegating
A client-side assessment view. Some clustering problems are manageable in-house. Others become good candidates for a specialist studio once geometry, density, drift, or outlier pressure start dominating the work.
Comfortable in-house
A competent internal team can often handle this without specialist help.
Possible, but assumption-sensitive
Can stay in-house if the team has time and the geometry is treated carefully.
Strong case to delegate
This is where specialist clustering work usually becomes worth the cost.
Wide / high-dim data
Uneven density
Drift
Mixed data
Noise / outliers
Standard in-house baseline
Good when the data is tidy, balanced, and close to the assumptions a default workflow expects.
rarely enough
poor fit
goes stale
poor fit
center gets pulled
Advanced in-house team
Possible in-house, but only if the team can spend real time on geometry, validation, and method selection.
possible
possible
if monitored
if designed well
if policy is clear
Good candidate to delegate
A strong case for an expert studio when the business consequence is high and the geometry is clearly hostile to defaults.
delegate
delegate
delegate
delegate
delegate
K-means forces spherical clusters onto non-spherical structure
Manifolds, elongated shapes, and uneven cluster sizes make the default centroid answer look tidy but wrong.
Outliers and variable density drag the centers around
Noise-heavy or non-homogeneous datasets often need density-first or graph-aware workflows instead of centroid baselines.
Drift makes last quarter’s clustering stale
Telemetry, operations, and recurring data feeds need refresh rules and monitoring, not one frozen set of labels forever.
How engagements begin
The service begins with the brief, turns that into a geometry and risk assessment, and then returns the plan of action. That keeps the work grounded in the decision the business needs to make instead of in whichever algorithm is easiest to run first.
1
Brief review
Review the brief and determine where clustering is actually being asked to carry the decision.
2
Assessment
Assess geometry, density, similarity choice, drift, and outlier behavior before approving a method family.
3
Plan of action
Return a plan of action that ends in either clustered outputs plus reporting, or an approved clustering API.
What the client receives
A clearer view of the deliverables: cluster assignments and confidence, diagnostics, interpretation, and the option of operational delivery when the methodology is ready to repeat.
Primary outcome
This is a core part of the delivery path.
Supporting role
Present, but not the main reason the client chooses that path.
Clustered outputs and reporting
Approved API delivery
Cluster assignments and confidence
Hard labels, soft probabilities, edge cases, and noise handling.
included
included
Validation and diagnostics
Why the approved method fits, where it fails, and what the business should trust.
included
reused
Interpretation and report
What the groups mean for routing, monitoring, prioritization, personalization, or investigation.
included
supports
Operational repeatability
How the same approved workflow is reused on future data.
future option
primary outcome
Dataset fit
Business problems that tend to escalate fast
Catalog and document grouping
Product catalogs, multilingual content, and search-result grouping where cosine and k-means often separate by artifact instead of by business intent.
Sensor and telemetry states
Manufacturing, fleet, and operational telemetry where drift, non-stationarity, and variable density make stale labels expensive.
Mixed operational records
Datasets that combine numeric, categorical, sparse, and text-derived features and need a deliberate similarity design rather than a default metric.
High-dimensional business data
Embeddings, wide records, and complex behavior profiles where distance concentration and non-flat geometry break naive clustering.
Case studies
The strongest proof is when the default approach breaks
Clients rarely call because clustering is impossible. They call because the easy answer became brittle, vague, or too risky to keep operating.
Manufacturing
anonymized engagementFailure-state discovery in drifting manufacturing telemetry
The client received clustered outputs, interpretable state definitions, and a production-ready path for re-running the approved methodology on future telemetry.
4
failure states
22%
downtime reduction
When to re-run
operating rule
Commerce and knowledge systems
anonymized engagementCatalog and multilingual document grouping under sparse similarity
The client received grouped outputs that aligned with business intent instead of language islands, plus a repeatable path for applying the approved methodology to future catalog and document batches.
Intent-led groups
delivery outcome
Less manual review
operational effect
API-ready workflow
repeat path
Catalog and multilingual grouping before and after the approved workflow
A harder business problem than simple stratification: sparse, multilingual records fragment under default similarity and only become usable after the right workflow is approved.
Language or taxonomy islands
Recovered intent groups
Noise or duplicate-heavy edge cases
Technical depth
The technical notes now lead with the business failure first.
Each deep dive starts from the business failure, explains what is breaking in plain terms, and then shows what changes when the workflow is fixed.
Read the deep dives
Start here
Send the brief. The first deliverable is the assessment and plan of action.
The goal is not to force a clustering model into every dataset. The goal is to determine what structure exists, what the business can trust, and whether the right outcome is clustered outputs, a report, or an approved API.
Start a technical review
Send the brief, get an assessment, and receive a plan of action within one business day.