Enterprise AI
Leadership, architecture and delivery at organisational scale.
Local AI first ≠Local AI only
Does every AI workload really need the cloud?
I experiment with real systems to find what should run Local, Hybrid or Cloud, proving trade-offs with reproducible benchmarks and code.
Principal AI Engineer & AI Strategy LeadExplore method & systems
Where should AI run?
Start with the workload, not the model. Compare what Local, Hybrid and Cloud make possible—and the constraints each one introduces.
THREE WAYS TO RUN AI
The right architecture depends on the workload.
Swipe to compare
- 01 / 03Local AION-DEVICE
Privacy and control, by design.
Run AI close to the data. Keep sensitive processing on the device and own the runtime end to end.
THE TRADE-OFFS
- Privacy
- Data stays on device
- Latency
- Can work offline
- Control
- Own the runtime
- Models
- Device-bound
- Scale
- Hardware-limited
BEST WHENSensitive workloads, offline reliability and predictable local interactions.
- 02 / 03HybridPOLICY-ROUTED
Control where it matters. Power when needed.
Keep critical steps local, then route selected tasks to remote capacity when the workload calls for it.
THE TRADE-OFFS
- Privacy
- Split by policy
- Latency
- Local fallback possible
- Control
- Shared across layers
- Models
- Route by workload
- Scale
- Cloud burst when needed
BEST WHENMixed workloads with different privacy, capability and capacity needs.
- 03 / 03CloudREMOTE INFRASTRUCTURE
Frontier capability. Elastic scale.
Use hosted models and managed infrastructure when reach, model strength and variable demand take priority.
THE TRADE-OFFS
- Privacy
- Remote processing
- Latency
- Network-dependent
- Control
- Provider-managed
- Models
- Frontier access
- Scale
- Elastic capacity
BEST WHENHigh-capability workloads, burst demand and centralized operations.
01 / 03
NEXT · BUILD
Explore the runtime layer Choosing where AI should run is the first step. Making that runtime reliable is the next.
01 DECIDE 02 BUILD
Build the reusable AI layer once.
When Local or Hybrid makes sense, I build reusable infrastructure for reasoning, mobile inference, repeated decisions and private speech.
Execution environments
Two platforms. Dedicated runtimes.
MAC · DESKTOP & SERVER01 / 04
Korgis
Local AI runtime control plane
Run and manage local AI models behind one stable application-facing API.
ANDROID · ON-DEVICE02 / 04
Harnex
Shared Android local-AI control plane
Run local GGUF models once and expose governed on-device inference to multiple apps.
Specialized components
Built for specific workloads
DecisioPrefill stable context once, send only changing state and return a typed action.experimentalLocal ASRTurn speech into text locally so audio does not have to leave the controlled workflow.active
04 reusable systems02 execution environments02 specialized components
Each system solves a different execution problem. Real product integrations are shown in the following TEST stage.
02 BUILD 03 TEST
Does it survive a real product?
Infrastructure only matters when real applications can depend on it. These products put privacy, usability and integration assumptions to the test.
REAL PRODUCTS · REAL CONSTRAINTS
What happens when the runtime meets a product?
Three applications, three different workload constraints.
Swipe to explore
RedactGuardWEB & DESKTOP · PII REDACTIONSensitive documents. Local processing.
Can sensitive document processing stay local?
PII redactionHuman reviewPrivacy boundary
Aura FinanceMOBILE · ON-DEVICE PARSINGUnderstand transactions on-device.
Can personal transactions be categorized on-device?
On-device parsingDeterministic mathZero telemetry
ClosedRoomMACOS · MEETING INTELLIGENCEMeeting intelligence. Private by default.
Can meeting intelligence stay inside the room?
Local audio ASRSpeaker contextOpt-in cloud
01 / 03
NEXT · MEASURE
Explore the evidence Real applications expose the limits. Measurement shows where to improve.
03 TEST 04 MEASURE
04 MEASURE
Test the assumptions, not just the demo.
Reproducible benchmarks turn architecture choices into evidence: what works, how well, and at what cost.
Benchmarks
SWIPE TO EXPLORE
01 / 06
Loading published evidence…
NO PUBLIC SNAPSHOT YET
Real runs exist outside the published repository snapshot. This view stays empty until a versioned evidence artifact is available.
PUBLISHED BENCHMARK SNAPSHOT
Source WHAT THE EVIDENCE SHOWSPublished result
RELATED EVIDENCE · MODEL CAPABILITY BENCHMARK
MCB source ↗JEV and GPT Luna, tested on the same task
Intent classification · comparing identical benchmark signatures
This is Model Capability Benchmark evidence, not a result from the separate Jev vs LLM suite.
Explore separate local-routing runsDifferent benchmark · includes latency measurements
WHAT THE RESULTS SHOW—
Accuracy difference—
Median response time—
ROUTING · EQUAL REQUEST COUNT
—How local models compare
Runs with the same request count are shown together; identical samples and configurations have not been independently verified.
PRIVACY BENCHMARK
Measured evidenceHow well do local models protect sensitive data?
MEASURED PERFORMANCE
Awaiting published runsWhat the results tell us
KEY TAKEAWAY
Compare the models
One run · same benchmark workload
Recall ↑ Leakage ↓
System recall includes inference failures. The p95 shown here was observed during quality runs, not a repeated latency benchmark.
The test dataset is published, but no measured model results are available for this dataset yet.
Explore deeper evidenceQuality × latency, cross-dataset matrix and methodology
SAME-RUN QUALITY × LATENCYPrivacy quality vs observed response time
System recall ↑ · observed p95 →Swipe across the chart for full-size labels
The quality–latency comparison requires at least two models measured in the same complete run.
ACROSS DATASETS
Model × dataset
Individual dataset scores, never a global rating.
— means that this model/dataset has no published measurement, not 0% performance.
Run comparisons are exploratory. Healthcare and Ambiguous have distinct model cohorts; do not interpret their figures as a controlled leaderboard. Dataset gold spans have not been independently human-reviewed.
AT A GLANCE
Two numbers that put the finding in context
Compare models and speedScores, latency and all measured models
Fastest local—
MODEL COMPARISON
Explore measured models
Compare published scores. Local and API latency are not hardware-matched.
SWIPE TO COMPARE
Performance chart & model inspectionQuality vs median latency from the published overview
QUALITY × LATENCYLocal routing frontier
Accuracy vs p50 latencySwipe across the chart to inspect all models
INSPECT MODEL—
—Measured modelPareto frontier
Methods & limitationsScope, caveats and source details
EVALUATION SCOPE Context, not performance
— models— requests— dataset— runtime
Aggregate quality is reported by the published Model Capability Benchmark. Latency is the reported median (p50); local and API runtimes have different execution environments.
These figures describe the published snapshot, not a controlled hardware-normalized leaderboard. Use the linked source for benchmark signatures, tasks and reproducibility details.
SOURCE UNAVAILABLE
Published evidence could not be loaded.
No fallback or invented performance values are shown.
Who I am
AI strategy. Hands-on engineering. Local AI.
I lead enterprise AI systems and build AI on real devices to understand what should run Local, Hybrid or Cloud.
Hands-on engineering
AI systems built and tested on Mac, Android and local infrastructure.
Public work
Open-source systems and technical writing through GitHub and stAI tuned.