Coming soon — early draft

Architecture Radar

A running list of the data and AI technologies I'd actually put in front of a client right now, organized the way I make the call myself: what I'd adopt without hesitation, what's worth a real trial on the next suitable project, what I'm still assessing before betting production on it, and what I'd steer a new project away from — hold. This page is a work in progress and will fill out over time; treat it as a snapshot of current judgment, not a finished catalog.

The radar at a glance

Four quadrants × four rings. Filled = recently added or moved, hollow = established. Click any blip to open the assessment or deep-dive.

Recently added / moved Established

Adopt

Proven, low-regret choices I default to unless there's a specific reason not to.

Apache Iceberg
The default open table format for a new lakehouse build — broad engine support, a real catalog ecosystem, and no single-vendor lock-in.
dbt
The default transformation layer wherever SQL-first modeling fits — testing, lineage, and documentation come for free instead of being bolted on later.
Kafka
Still the default event backbone for anything that needs a durable, replayable log — the operational maturity is hard to match.
Snowflake / Databricks as the core warehouse-lakehouse
Either is a safe default for the platform of record, depending on the org's existing cloud and team skill set.

Trial

Ready for a real project, not yet a blanket default.

StarRocks / Apache Doris — read the full assessment →
Genuinely fast for sub-second, high-concurrency serving — worth it when that's the actual bottleneck, not a default replacement for your warehouse.
ClickHouse — read the full assessment →
The fastest, cheapest analytical scans in class for observability, events, and customer-facing dashboards — adopt for that niche, not as a general warehouse.
MCP for agent tool integration
The right default for exposing tools to an agent when the connection is genuinely external and needs a real security boundary.
Agent Skills (progressive disclosure)
A strong pattern for packaging procedural knowledge without bloating context — still less standardized than MCP, worth piloting deliberately.
Fabric Data Agent / Foundry Agent Service
The delegated-identity security model is the right idea; still maturing operationally enough to want a scoped pilot before a wide rollout.

Assess

Watching closely — not yet betting a production system on these.

Meta-harness platforms (e.g. Omnigent)
Cross-agent governance and cost control is a real gap worth solving — the composition story is still ahead of its production track record.
The Ralph-loop style of agentic automation
A genuinely clever pattern when paired with real verification grounding — still task-shaped, not yet a general default.
Lakehouse-native serving over Iceberg (no ETL copy)
The direction the whole stack is heading — worth designing new pipelines around, with fallback plans while tooling catches up.

Hold

I'd actively steer a new project away from starting here.

New on-prem Hadoop/Cloudera builds
The ecosystem has moved on — there's no good reason to start a greenfield project here in 2026.
Fully autonomous agent loops with no verification grounding
The AutoGPT-era pattern — impressive demos, unreliable in production without a real check-the-work step.
Hand-rolled ETL for problems dbt/Dataflow-class tools already solve
Custom orchestration code for standard transformation and windowing logic is technical debt on day one.