No shared truth
Listings, images, free text, long documents, partner feeds, and private records described the same physical assets in different ways.
Client work / Property intelligence
From fragmented property data to a production AI platform.
Over six months, Drizzle built a production platform that now processes approximately one million data events daily across self-hosted models, training pipelines, and dynamically routed LLM services.
ENGAGEMENT RECORD / PRODUCTION SYSTEM Shared intelligence core
Market feeds Images Documents Private data
01Self-hosted ML
02Self-hosted LLM
03OpenAI routing
04Anthropic routing
CANONICAL CONTRACT Property intelligence
Assets, observations, evidence, lineage, and predictions
Valuation Document review Research Workflows
1 million daily events 10 training pipelines GPU accelerated
PRODUCTION SCALE 1 million Data events processed daily
MODEL PORTFOLIO Around 20 Self-hosted ML and LLM models
TRAINING SYSTEM 10 Production training pipelines
GPU DEPLOYMENT 30x Observed acceleration against the earlier execution path
The constraint
The hard problem was fragmented truth, not model availability.
The company had valuable data and clear product opportunities. What it lacked was a shared representation and an operated path from raw evidence to a trustworthy product decision.
Repeated product plumbing
Every new product risked rebuilding ingestion, identity, feature, persistence, and model-serving logic before it could create customer value.
Low operational trust
Model outputs lacked a consistent contract for evidence, uncertainty, lineage, review, deployment, and recovery.
THE PLATFORM DECISION
Build one canonical intelligence core, then let customer products
specialize at the edges.
Six-month engagement
Strategy, implementation, and operating reality stayed connected.
Drizzle worked across product architecture, machine learning, data, serving, security, delivery, and production operations as one engagement rather than a chain of disconnected projects.
-
Define the product boundary
Drizzle aligned product, data, model, platform, and customer context around one durable domain contract and a staged delivery plan.
-
Create the shared intelligence core
We implemented canonical property identity, multimodal enrichment, purpose-built stores, and stable service contracts.
-
Productionize both execution paths
Real-time and scheduled workloads reused the same model semantics while receiving independent scaling, placement, and reliability profiles.
-
Prove and hand over the system
GitOps, security, observability, quality gates, recovery controls, runbooks, and runtime checks made ownership transferable.
DELIVERY TEAM Approximately 10 engineers
The engagement connected the disciplines required to turn product intent into a system that could be trained, deployed, operated, and evolved in production.
- 01 MLOps
- 02 Data engineering
- 03 Full-stack engineering
- 04 Platform and SRE
- 05 AI and LLM engineering
- 06 Data science
The system
One domain contract connected evidence to product action.
Stable boundaries kept data, models, infrastructure, and products independently evolvable while a routing layer selected between self-hosted models and major hosted LLM APIs for each workload.
01 / INGEST
Market feeds User uploads Partner data Long documents
02 / MODEL AND ROUTE
Around 20 self-hosted models Dynamic OpenAI and Anthropic routing 10 training pipelines Real-time and batch serving
03 / CANONICALIZE Shared property intelligence
Identity, observations, embeddings, relationships, predictions, uncertainty, evidence, and lineage.
04 / ACTIVATE
Valuation and triage Document review Market research Private workflows
OPERATE Kubernetes serving Workflow orchestration GitOps delivery Observability Identity and secrets Backup and recovery
What Drizzle delivered
The models were only one layer of the engagement.
The deliverable was a reusable product and operating foundation, not a collection of notebooks or a platform diagram left for another team to implement.
Canonical intelligence core
Versioned assets, observations, relationships, predictions, evidence, quality signals, and lineage became the shared product contract.
Model portfolio and routing
Around twenty self-hosted ML and LLM models worked alongside a chat layer that dynamically routed requests across OpenAI and Anthropic APIs.
Serving and training paths
Real-time services, scheduled enrichment, and ten production training pipelines reused governed data and model contracts.
Trust-bearing product APIs
Predictions travelled with uncertainty, explanations, source evidence, missingness, schema versions, and model versions.
Production platform foundation
Kubernetes serving, orchestration, GitOps, identity, secrets, policy, telemetry, backup, and recovery formed one operating substrate.
AI software factory
Cursor, Codex, Claude Code, and OpenCode operated through engineered harnesses, delivery loops, and evidence-based AI code verification.
AI software factory
Agent speed was paired with engineered verification.
The delivery system coordinated Cursor, Codex, Claude Code, and OpenCode through advanced harness engineering and loop engineering. Shipmoor added evidence-based verification before agent-generated changes could be treated as complete.
Explore the verification layer Cursor Codex Claude Code OpenCode
- 01Harness engineering
- 02Loop engineering
- 03AI code verification
VERIFICATION / SHIPMOOR Evidence-backed change
Measured outcomes
Credible proof is bounded, labeled, and reproducible.
Production scope, implemented capability, runtime observations, and validation results are shown with their context. Drizzle does not turn an internal signal into a universal performance promise.
Data events supported and processed by the production platform across ingestion, enrichment, persistence, and product workflows.
Self-hosted ML and LLM models operating behind shared serving, observability, security, and delivery patterns.
Production training pipelines using governed data, artifact, evaluation, and promotion paths.
Performance improvement produced by service architecture, request-contract, execution, and model-serving changes against the earlier baseline.
Acceleration achieved on the GPU deployment path against its previous execution baseline.
Strict platform and model-serving smoke checks passed at a major migration checkpoint.
Performance scope. The 10x architecture gain and the 30x GPU deployment gain compare different execution paths with their respective earlier baselines. They are distinct observations and are not multiplied together.
Product impact
Platform primitives became decisions users could trust.
Shared intelligence shortened the path from a new use case to a safe, explainable customer workflow while private context stayed at the product edge.
Canonical market intelligence
- Before
- Repeated source plumbing
- After
- One queryable asset identity with observations, embeddings, history, quality, and predictions.
Explainable valuation
- Before
- A number without context
- After
- Purchase and rent estimates with calibrated bands, domain guidance, and signed feature contributions.
Evidence-grounded document AI
- Before
- Manual long-form review
- After
- Structured findings from irregular reports with page references and deterministic recovery.
Customer-specific workflows
- Before
- A forked AI stack for every customer
- After
- Private data, policy, evidence, and integrations layered over reusable intelligence APIs.
Ownership at handoff
The client received a system it could extend, not a dependency on an agency.
Drizzle stayed close enough to production to make the platform operable, while designing every boundary for long-term client ownership.
- 01 Platform and application code
The implementation remained in the client's repositories and infrastructure boundary.
- 02 Domain and service contracts
Versioned schemas and APIs let models and runtime components evolve without forcing product rewrites.
- 03 Operational evidence
Dashboards, alerts, smoke checks, quality gates, backups, and recovery paths exposed the real state of the system.
- 04 Engineering operating model
Decision records, GitOps flow, AI agent harnesses, Shipmoor verification, runbooks, and handoff practices made future delivery repeatable.
Your platform constraint
Bring the fragmented system. We will help make the production path coherent.
Start with the workload, the data boundary, and the operating reality. We will identify the first platform decision worth proving.
Talk to an engineer