Case studies

Work we'd defend in a deposition.

Each case study is signed off by the client. Names changed only where contracts require it.

FINANCIAL SERVICES · KYC

First Citizen Bank — KYC reasoning agent

An evidence-grounded agent reading 11 source systems to produce analyst-ready KYC drafts, gated by human approval.

−62%
Manual review time
+18pp
Analyst throughput
99.4%
Eval precision

CHALLENGE

Analysts spent 47% of cycle time on document collation rather than judgement. KYC quality varied team-to-team.

APPROACH

RAG over onboarding docs + Claude reasoning agent with structured tool-use into core banking systems. Human approval gate before any analyst-visible publish.

STACK

ClaudepgvectorBedrockTemporalDataDogRagas
RETAIL · CATALOG

Retail — Catalog enrichment agent

An agent enriching 24M SKUs with structured attributes, accurate categorisation, and shopper-friendly descriptions across 14 locales.

24M
SKUs enriched
+18%
AOV after rollout
−71%
Catalog onboarding time

CHALLENGE

Inconsistent attribution and copy across suppliers blocked search and merchandising. Manual catalog ops capped expansion.

APPROACH

Multi-step agent (extract → classify → translate → enrich) with golden-set eval per category and continuous reviewer feedback loop.

STACK

GPT-4oCohere rerankAzure OpenAISnowflakedbt
SAAS · OBSERVABILITY

Observability copilot

A Slack-native engineering copilot that triages alerts, surfaces runbooks, and drafts first-hypothesis comments inside the on-call channel.

−47%
MTTR (first hypothesis)
1,400
Daily active engineers
94%
Helpful-rated responses

CHALLENGE

Knowledge fragmented across runbooks, post-mortems, code, and tribal Slack threads. New on-call engineers paid a long ramp tax.

APPROACH

RAG over Git + Confluence + Slack archive, ranked by recency × relevance. Slack-native UX. Per-team eval golden sets owned by SREs.

STACK

ClaudeQdrantGCPSlack APILangfuse
HEALTHCARE · PRIOR-AUTH

Health — Prior-auth automation

Document-driven workflow extracting clinical evidence, matching policy, and drafting prior-auth requests with clinician review gates.

−58%
Time to submission
+22%
First-pass approvals
100%
Clinician-reviewed

CHALLENGE

Prior-auth backlog delayed patient care. Each request required manual cross-referencing across EHR notes, lab results, and payer policy.

APPROACH

HIPAA-aware pipeline designed for clinical workloads: PII redaction → evidence extraction → policy-matching agent → clinician approval → payer submission. Audit trail per packet.

STACK

ClaudeAzure OpenAIFHIRpgvectorTemporal

Aggregate outcomes

Across 40+ shipped engagements.

94%
Engagements reaching production
11 wks
Median time to first production user
3.8x
Median ROI within 12 months

Want the long-form, signed version?

Detailed case studies (architecture, eval results, references) available under NDA for serious buyers.

Request the case study pack