codeAtelier
Menu

FLAGSHIP SPECIALTY · ENTITY GRAPHS & MDM ALTERNATIVE

Reltio-grade entity graphs — AI-native, evidence-backed, and 100% yours.

Escape $150k–$500k+/yr MDM SaaS rent and brittle match-rule sprawl. We engineer a client-owned entity resolution system inside your cloud via a one-time build and corpus index, then run incremental deltas for pennies on the dollar.

Systems Architecture

Five engineered stages — not a naive LLM prompt wrapper.

Prompting a frontier model row-by-row fails on cost, determinism, and auditability. Production entity resolution treats language models as bounded reasoning engines inside a strict indexing, caching, and governance control plane.

01

Indexed Recall First

Vector, lexical, and geo blocking narrows millions of raw records into tight, high-recall candidate clusters before an agent ever runs.

Agents never brute-force O(n²) comparisons across your warehouse. Hybrid indexes isolate only plausible candidate neighborhoods so expensive reasoning runs strictly on genuine edge cases.

02

Agentic Evidence Research & Cached Evidence Library

Ambiguous records are researched and resolved once, then persisted as cited, content-hashed evidence snapshots.

Every merge or link cites the exact registry extract, domain match, or internal feed that proved it. Subsequent delta runs reuse cached proof at near-$0 instead of re-paying for the same research.

03

Corporate Hierarchy, Brand & Site Resolution

Unifies Legal Entity ↔ DBA / Operating Site ↔ Brand / Franchise Flag ↔ Ultimate Parent across messy real-world records.

Legacy fuzzy string rules break when a single legal entity operates under multiple trade names, regional branches, or post-acquisition rebrands. Evidence-backed graph edges model ownership and operational sites distinctly.

04

Governance & Safety Control Plane

Hard per-run and per-entity spend caps, observer audit loops, 3-band confidence routing, and safe unmerge/split.

Auto-resolve high-confidence matches, route borderline cases to a human steward queue, and reject noise. Every edge stores clickable provenance for deterministic graph replay and safe unmerge/split without breaking downstream keys.

05

Incremental Delta Sync

Record fingerprinting isolates only net-new or mutated records so steady-state refreshes cost pennies on the dollar.

Unchanged entities skip inference and research completely ($0.00). Downstream warehouses and operational apps receive clean, versioned golden IDs and lineage diffs on every scheduled or event-driven sync.

Architectural Tradeoffs

How we compare — and when to pick someone else.

Every resolution architecture has an honest sweet spot. Here is how a client-owned AI-native graph stacks up against legacy MDM suites and embedded or open-source matchers.

Dimension Legacy MDM SaaS Reltio, Informatica, Profisee, Semarchy Embedded / OSS Matchers Senzing, Tilores, Splink, Zingg Client-Owned AI-Native Graph codeAtelier Architecture
Pricing & Licensing Model $150k–$500k+/yr recurring SaaS rent, tiered by record volume, plus multi-quarter SI fees and annual uplift. Low software license or open-source core, but custom glue code, hosting, and research connectors sit on your team. Fixed-scope build + 1-time initial corpus index. Zero recurring vendor license rent; you pay only raw VPC compute & token usage.
External Web / Registry Evidence Research Relies on expensive third-party D&B / data-broker add-ons or manual steward web lookups outside the tool. None out of the box; matches strictly across attributes already present in your input tables. Built-in agentic evidence gathering (air-gapped licensed feeds or sanitized public business lookups) with a content-hashed evidence cache.
Corporate Hierarchies, Rebrands & DBAs Rigid hierarchy tables and hand-tuned match rules that degrade rapidly across acquisitions, DBAs, and franchise structures. Flat clustering or pairwise record linkage; struggles when names share zero lexical overlap after a rebrand. Explicit graph modeling across Legal Entity, Operating Site / DBA, Brand / Franchise Flag, and Ultimate Parent backed by cited proof.
Periodic Delta Economics Annual license resets regardless of low change rate; re-enrichment add-ons charge per record refreshed. Cheap CPU compute on deltas, though often requires full re-blocking or manual threshold retuning as data drifts. Record fingerprinting + cached evidence library: unchanged entities cost $0.00; only net-new or mutated records trigger delta work.
Data, Code & IP Ownership Proprietary vendor black box; schemas, match tuning, and survivorship logic remain trapped inside vendor SaaS. High ownership for pure OSS (Splink/Zingg); partial vendor lock-in for commercial embedded matchers. 100% client-owned repository, Terraform/IaC, graph schema, evidence store, and steward UI running in your own cloud account.
VPC & PII Isolation Boundary Data replicated to multi-tenant vendor cloud; requires DPAs, vendor security reviews, and cross-boundary sync. Runs inside your cloud or warehouse, though you build the egress controls, audit trails, and LLM governance yourself. Runs natively inside your AWS, Azure, or GCP VPC with private LLM endpoints, strict PII scrubbing, and zero studio data touch.
When to pick them instead of us An already-staffed 15-person Informatica steward organization running tightly standardized banking or regulatory golden-record workflows. Clean <50k single-country internal records with zero external research or corporate hierarchy needs — run SQL or Splink in-house. Messy B2B/vendor/site graphs (100k–10M+ records), looming $150k+ MDM renewals, or custom hierarchy & provenance requirements.

Transparent Economics · ~500k Entities

Honest 3-Year Full TCO (36-Month Illustration)

Illustrative 36-month total cost of ownership for an enterprise corpus of roughly 500,000 entities with ongoing monthly deltas, hierarchy tracking, and stewardship workflows.

Cost Component (36 Months) Legacy MDM SaaS + SI Rollout codeAtelier In-VPC Entity Graph
Initial Implementation & Corpus Index $200k–$450k Multi-quarter SI discovery, rule authoring, and data-broker connectors $75k–$115k 1-time build, recall indexes, steward UI, and full 500k initial graph index
36-Month Platform License Rent $450k–$1.5M+ $150k–$500k+/yr recurring SaaS tiers + annual renewal escalation $0 Zero recurring vendor license rent; 100% client-owned codebase & schema
36-Month Delta Compute & LLM Tokens $60k–$180k Third-party data-broker match/enrichment fees & overage tiers $5k–$15k Fingerprinted deltas + cached evidence reuse; unchanged entities cost $0.00
Ongoing Operations & Upkeep (3 Yrs) $90k–$240k Dedicated MDM admin / rule-tuning contractor retainers $30k–$50k Light internal engineering upkeep or optional fractional principal advisory
Estimated 3-Year Full TCO & Asset Ownership ~$750k–$1.5M+ Zero owned resolution engine at Year 3; renewal lock-in continues ~$110k–$180k 100% owned IP, full audit trail, and $0 recurring SaaS rent

Security & Isolation Boundary

Client-VPC execution & strict PII egress specification.

Your entity graph deploys directly inside your own AWS, Azure, or GCP VPC. Zero source data or resolved graphs ever touch codeAtelier servers. Pick the egress posture your security team mandates:

Mode A · Zero Outbound Internet

100% Air-Gapped VPC Mode

Everything executes strictly inside your private network boundary over internal warehouse tables and your existing licensed feeds. Inference routes solely through private VPC LLM endpoints (AWS Bedrock, Azure OpenAI, or GCP Vertex AI) with zero external web calls.

Mode B · Allowlisted Corporate Verification

Sanitized Public Research Mode

When resolving ambiguous DBAs, subsidiaries, or corporate rebrands, an egress sanitizer strips all sensitive fields before an allowlisted lookup runs. Only sanitized public business keys (company name, domain, street address) leave the VPC — zero customer PII, personal names, financials, or internal IDs, and zero data ever touches studio servers.

Technical FAQ

Architecture, cost, and safety questions answered directly.

How is an AI-native entity graph different from Reltio or Informatica MDM?

Legacy MDM suites charge $150k to $500k+ every year for proprietary matching engines powered by brittle deterministic rules and expensive third-party data add-ons. Our approach builds a client-owned entity graph inside your own VPC. Indexed blocking narrows candidates first, then AI agents resolve ambiguous corporate hierarchies, DBAs, and rebrands using verifiable evidence snapshots — eliminating recurring vendor license rent while giving your team full control of code, schema, and provenance.

Why doesn't agentic AI entity resolution blow up token costs on 500k+ records?

We never send raw tables into an LLM or run O(n²) pairwise prompts. Vector, lexical, and geo blocking compresses 500k+ records into compact candidate clusters, and deterministic matching settles exact overlaps at zero token cost. Agent reasoning runs only on genuine ambiguity, backed by a content-hashed cached evidence library and hard per-run spend caps. After the initial index, incremental delta runs fingerprint mutations so unchanged entities cost $0.00.

How do you prevent hallucinated merges and support safe unmerging in production?

Every graph decision passes through a strict governance control plane: automated observer audit loops verify claims against cited evidence snapshots, and three-band confidence routing separates auto-merges from human steward review queues and rejections. Every edge records immutable provenance for deterministic graph replay. If new evidence arrives later, our split workflow performs a safe unmerge without corrupting downstream golden IDs or breaking dependent warehouse tables.

Can this run inside our AWS, Azure, or GCP VPC without leaking customer PII?

Yes. All pipelines, graph storage, and steward UIs deploy directly into your AWS, Azure, or GCP account with zero data touching studio servers. You choose between 100% Air-Gapped Mode — using internal feeds and private VPC LLM endpoints like Bedrock, Azure OpenAI, or Vertex — or Sanitized Public Research Mode, where allowlisted egress sends only public business lookup keys such as company name, domain, and street address, never customer PII.

How fast can we validate this on our own data before committing to a full build?

We start with a focused diagnostic sprint on a representative slice of your toughest entity records — typically two to three weeks inside your environment or on a de-identified sample. You receive measured precision and recall against your ground truth, verified hierarchy accuracy, actual token and compute telemetry for your corpus scale, and a fixed-scope implementation plan before committing to a full production build.

Direct Principal Intake

Evaluating an MDM renewal or drowning in unresolved corporate hierarchies?

Send 5 lines to hello@codeateliervb.com. A principal engineer replies within one business day with a candid assessment — including if you are better off keeping SQL or Splink in-house.

  • 01Source systems & approximate record scale
  • 02What is breaking today or coming up for renewal
  • 03Cloud, warehouse, and private LLM constraints
  • 04Target decision or cutover timeline

Direct Engineering Channel

hello@codeateliervb.com

US-East & European business hours. Direct US/EU B2B contracting (USD or EUR, W-8BEN-E ready) or embedded alongside your preferred systems integrator.