Kuzushi Labs · Application library
FIELD in practice
Thirty-six worked applications, each following the customer outcome through investigation, engineering, evidence and ownership.
The figures are stated planning assumptions, calculated implications and proposed acceptance thresholds. These analyses demonstrate the method and do not report client measurements. Replace their inputs with evidence from your own engagement.
36 of 36 applications
Customer delivery · 4 min
Choosing the first useful customer workflow
Compare three candidate workflows; deliver one permitted journey.
Customer delivery · 7 min
Enterprise search that respects changing access
One governed corpus and identity provider; answer with source references.
Customer delivery · 4 min
A legacy connector with an honest recovery plan
Initial backfill, incremental sync, deletion and reconciliation for one source.
Customer delivery · 4 min
An assistant inside the customer network
Read-only SQL and ticket lookup through the approved outbound proxy.
Customer delivery · 4 min
Customer support with controlled refunds
One refund policy and payment integration, with explicit approvals.
Customer delivery · 4 min
Recovering a pilot whose users are leaving
Diagnose and improve one user journey before extending features.
Customer delivery · 4 min
Protecting customers during a quota shock
Fair admission, bounded queues and an approved degraded service policy.
Customer delivery · 4 min
Moving a service into customer ownership
Transfer deployment, credentials, recovery, monitoring and support ownership.
Agents and workflows · 4 min
Financial research with inspectable evidence
One document corpus and defined issuer, currency, period and as-of semantics.
Agents and workflows · 4 min
Natural-language analytics with governed metrics
Approved semantic metrics, read-only views and one analyst workflow.
Agents and workflows · 4 min
Investigations that survive waiting and restarts
Five read-only evidence integrations and a durable approval boundary.
Agents and workflows · 4 min
Useful memory without lost obligations
Typed memory, correction, deletion and authoritative task obligations.
Agents and workflows · 4 min
When agent delegation earns its cost
Compare a fixed workflow, one bounded agent and specialist delegation.
Agents and workflows · 4 min
From a repository issue to a reviewed change
Issue intake, isolated patching, tests and human review for one repository.
Agents and workflows · 4 min
Evidence for operational anomaly investigation
Read-only access to approved logs, metrics and deployment events.
Agents and workflows · 7 min
A multilingual voice workflow with a staffed fallback
One caller journey in English and two Southeast Asian languages.
Models and data · 4 min
Audit the financial model before another training run
Audit templates, source overlap, training examples and agent behavior.
Models and data · 4 min
Improve tool use with a small, inspected dataset
One tool schema and error taxonomy, with supervised trajectory comparison.
Models and data · 4 min
Align service behavior without rewarding confident errors
Adjudicated preferences grounded in correct and permitted outcomes.
Models and data · 4 min
Ranking local services under live constraints
One market, defined relevance objective and current availability constraints.
Models and data · 4 min
Fraud decisions with delayed evidence
One transaction family, mature-label evaluation and explicit fallback policy.
Models and data · 4 min
Evaluate a marketplace change without misleading attribution
One policy change with a justified cluster or time-based assignment.
Models and data · 4 min
Language coverage that survives beyond the average
One domain, locale-aware examples and equal reporting by language.
Models and data · 4 min
Reproduce the path from training to serving
Version data, tokenizer, templates, model, adapters and runtime together.
Capacity and platforms · 4 min
Plan model capacity from real token demand
Serve the defined prompt/output distribution with admission limits.
Capacity and platforms · 4 min
Find the bottleneck behind an apparently idle CPU
Diagnose one service with request tracing, profiling and dependency timing.
Capacity and platforms · 4 min
Share GPU capacity without breaking interactive work
Separate admission and queues with explicit workload priorities.
Capacity and platforms · 4 min
Fit the workload, not just the model weights
Compare context limits, concurrency and quantization for one quality target.
Capacity and platforms · 4 min
Make the training run recoverable and economically bounded
One reproducible run with measured throughput and restart checkpoints.
Capacity and platforms · 4 min
Update an embedding index while serving customers
Versioned backfill, incremental catch-up and reversible query routing.
Capacity and platforms · 4 min
Recover durable work after a regional outage
Restore task state, identity, artifacts and external-effect reconciliation.
Capacity and platforms · 4 min
Prove a shared capability with a second customer team
One reusable contract, adapters and a supported integration path.
Agent infrastructure · 9 min
An invoice operations agent with controlled tools
One correction workflow with approved MCP tools, receipts and reconciliation.
Agent infrastructure · 6 min
Coordinate remote specialists without losing accountability
An A2A coordinator and Google ADK specialist with durable task records.
Agent infrastructure · 5 min
Operate coding agents across editors and isolated workspaces
One editor/client protocol path, sandbox runtime and reviewed patch workflow.
Agent infrastructure · 6 min
A shared agent service with tenant-level accountability
Versioned control plane, tenant budgets, durable runs and protocol adapters.