Agentry
Book a scope call
Use cases

Systems we’ve shipped. Patterns we build for you.

Every entry below ran in production with real users — clinicians, analysts, operators — and each one is a pattern we can adapt to your product or workflow. Filter by the kind of business you run.

/01SaaS · AnalyticsMulti-agent

Dashboards from a sentence — no SQL, no analyst queue

shipped at · Anlytic, multi-agent analytics platform
Anthropic ClaudeVercel AI SDKZod → JSON SchemaAgentic RAGSSE streaming

The problem

Business users with no SQL or BI background needed charts and answers every day — and every request queued behind an analyst. Routine reporting was the bottleneck for the whole company.

What was built

An agent layer inside the product. A cheap Haiku intent classifier routes each request to a specialist agent — chart generation, dashboard editing, data management, analytics Q&A — so large models are reserved for generation. Agents produce the full chart config and data query as schema-constrained structured output, so every chart is valid by construction. Analytics Q&A uses agentic RAG: instead of a stale vector index, agents retrieve live catalogs, table schemas and dashboard state through tools at query time, so answers reflect what the data says right now.

8chart types generated end to end
Dailyuse by non-technical business teams
0invalid charts — schema-checked before render
Livedata, not a stale index

For you

Any product that holds data — SaaS analytics, finance tooling, e-commerce, ops platforms — can get a plain-English layer that builds, edits and explains. We’ve done the hard parts already.

/02Professional servicesKnowledge

Trustworthy answers from company documents — with citations

built on · AWS Bedrock + Bedrock Knowledge Bases
Retrieve-then-ConverseNova · Claude · LlamaCitationsAbstention

The problem

Teams dig through contracts, policies and manuals for answers that are already written down. Generic AI chat gives fluent answers with no way to check them — a non-starter for anything a client or regulator might ask about.

What was built

A document Q&A assistant where every answer cites its exact source passage for auditability, and where the system says “not covered” rather than inventing an answer when the documents don’t contain one. Built on AWS Bedrock with Knowledge Bases (Retrieve-then-Converse) and model choice across Nova, Claude and Llama depending on cost and quality needs.

Secondsto a cited answer instead of a file hunt
100%of answers link their source

For you

Legal, accounting, consultancies, HR, customer support: point it at your document set, decide who can ask what, and roll out with the confidence that it won’t make things up.

/03OperationsAutomation

An AI executive assistant across email, Slack, WhatsApp and calendar

shipped at · Autonomix, “Donna” — Denmark
Claude Agent SDKElevenLabs voiceDjango · Celerypgvector hybrid searchNeo4j

The problem

Founders and operators lose their week to inbox, scheduling and follow-ups spread across four apps. Existing “assistants” asked people to move into a new tool — so nobody did.

What was built

A multimodal executive assistant with 76 tools across 9 business domains that works inside the apps people already use — email, Slack, WhatsApp, calendar — with voice and avatar front-ends. A progressive memory system condenses raw activity into curated profiles so per-request memory cost stays nearly flat as history grows. Agentic RAG over cross-platform communications blends vector, keyword, graph and recency signals. Relationship intelligence (entity resolution, RFM scoring, personality profiling) lets it prioritise the right contacts and adapt tone per relationship.

76tools across 9 business domains
Flatper-request memory cost as history grows
4channels — email, Slack, WhatsApp, calendar
Per-contacttone and priority

For you

The same architecture scales down to one workflow: intake and triage, meeting scheduling, follow-up drafting, CRM hygiene — with approvals on anything that leaves the building.

/04HealthcareKnowledge

A research assistant over millions of medical papers

shipped at · IQVIA (Fortune 500), clinical decision support
Azure OpenAI GPT-4PineconeRAG.NET 8 microservicesAWS EKS

The problem

Practising clinicians spent hours searching literature for treatment evidence and interaction risks — time that came straight out of patient care.

What was built

A RAG pipeline over millions of medical papers and clinical studies, shaped around how clinicians actually search rather than how documents are stored. On top of it, a clinical decision support system surfacing evidence-based treatment recommendations and drug-interaction warnings inside the clinicians’ daily workflow — designed with the end users, not delivered over the wall. Backed by .NET 8 microservices (Clean Architecture, CQRS, DDD) serving millions of patient records on AWS EKS.

~40%less research time for clinicians
Millionsof papers and patient records

For you

Clinics, health-tech, pharma services, insurers — anywhere the answer lives in a large, regulated corpus and the user has no time to browse it.

/05Any AI productCost

Cutting LLM spend 60–70% without touching quality

shipped at · Anlytic, in production
Context engineeringModel routingPrompt cachingClaude Haiku → Sonnet/Opus

The problem

Per-user serving cost was climbing with usage. The agent was dragging full tool schemas and every stale tool result through every turn, and using a large model for work a small one could do.

What was built

Context engineering as a discipline: tool-schema descriptions stripped before dispatch, stale tool-call results summarised out of message history, a cheap classifier routing requests so large models are reserved for generation, and prompt caching for the stable prefix. Cost per request tracked per user so the business could see the effect.

60–70%lower LLM API spend
Sameeval pass rate — quality held

For you

If your AI feature’s bill grows faster than its revenue, this is usually a 1–2 week engagement with a measurable before/after.

/06EngineeringQuality

A release gate so agent regressions never reach customers

shipped at · Anlytic, CI pipeline
evaliteHermetic harnessIn-process API fakeGitHub Actions

The problem

A prompt tweak or a model update could silently change agent behaviour. The first people to notice were customers, and the ticket said “the AI did something weird”.

What was built

A hermetic eval harness that runs 350+ graded scenarios against an in-process fake of the production API — no network, no flakiness. Each scenario scores task completion, hallucination, tool-call trajectory and step efficiency, and the suite collapses into a single CI pass/fail. Prompt and model regressions are caught before merge.

350+graded scenarios on every change
1CI pass/fail — no judgement calls at release

For you

Any team with an LLM feature in production. We build the harness around your real failure cases and hand it over — it becomes the definition of “working”.

/07EngineeringIntegration

Turning the APIs you already have into agent tools — in hours

shipped at · Anlytic, Amazon Bedrock AgentCore Gateway
AgentCore GatewayMCPREST · LambdaNo rewrite

The problem

Every new agent capability meant hand-writing and maintaining another tool wrapper. Onboarding a capability took days, and the tool layer became its own codebase to babysit.

What was built

Migrated the hand-rolled tool layer onto AgentCore Gateway, exposing existing REST endpoints and Lambda functions as MCP tools without rewriting them. Customer requests now turn into shipped capabilities in the same week they’re raised.

Days → hoursto onboard a new capability
0endpoints rewritten

For you

If you have an API, you have agent tools. This is the fastest route to an agent that can actually do things in your product.

/08Regulated dataSecurity

Agents that can’t be talked into anything

shipped at · Anlytic, enterprise accounts
Cedar policiesAgentCore Runtime · MemoryOpenTelemetryHuman-in-the-loop

The problem

Prompt-level guardrails (“never delete data”) are negotiable — a determined user or a poisoned document can talk a model out of them. Enterprise security teams, rightly, wouldn’t sign off.

What was built

Prompt-level guardrails replaced with Cedar authorisation policies enforced at the gateway, so even a hijacked agent cannot call tools outside its permissions — security enforced deterministically in infrastructure, not negotiated with the model. Agents moved to per-session isolated runtimes with durable memory so long-running jobs outlive request timeouts. Every run traced end to end (tool calls and intermediate reasoning), and human-in-the-loop approval added for schema changes, row edits and deletions: agents stage changes for review before execution.

Deterministicauthorisation — not prompt-dependent
100%of runs traced and diagnosable
Stagedwrites — a person approves risky actions
Adoptedby data-sensitive enterprise accounts

For you

The package that gets an AI rollout past security review in finance, healthcare and legal — or any company that handles customer data.

/09HealthcareData quality

Anomaly detection in regulated data pipelines

shipped at · IQVIA, clinical-trial data
ML anomaly detectionSnowflake.NET 8Kubernetes

The problem

Clinical-trial data flowed through pipelines where errors were caught late, by hand, under regulatory pressure. Manual review was slow and still missed things.

What was built

ML-based anomaly detection embedded in the data pipelines, flagging suspect records before they reached downstream systems, with review queues for the cases that needed a human.

+35%data accuracy
Lessmanual review for regulatory compliance

For you

Finance ops, logistics, insurance, health data — anywhere bad records are expensive and reviewers are scarce.

/10EnterpriseFoundation

A business-metrics platform that paid for itself many times over

shipped at · Selise, for European enterprise clients
AngularASP.NETMongoDBRabbitMQAzure

The problem

Enterprise clients needed configurable business metrics across heterogeneous data stores — and the messaging backbone couldn’t keep up with peak load.

What was built

A configurable business-metrics dashboard built from scratch against client requirements and data, plus a multi-level-query Hub System on microservices for heterogeneous storage. RabbitMQ tuned to halve messaging latency and raise peak throughput 12×. This is the kind of foundation we now put agents on top of.

~$2Mrevenue generated by the product
+10%annual revenue from the Hub System
12×peak messaging throughput
½the latency

For you

Proof that we build the boring, load-bearing parts properly — not just the model calls.

/11Field servicesMobile

An offline-capable field app for low-connectivity regions

shipped at · IQVIA, field data collection
iOSDuckDB · SQLiteSnowflake OLAPAWS

The problem

Field teams worked where connectivity was unreliable. Cloud-first tooling assumed the ideal case and failed in the real one — and large clinical datasets were expensive to query.

What was built

An offline-capable iOS field solution on DuckDB/SQLite that syncs when a connection appears, integrated with Snowflake OLAP for large datasets — meeting the deployment constraints of the field rather than the ideal case, and cutting AWS costs in the process.

Offline-firstworks with no signal, syncs later
LowerAWS costs on large datasets

For you

Construction, logistics, inspections, community health — any team whose “office” has no Wi-Fi.

Your workflow next

Which of these looks like your problem?

Tell us which pattern is closest and what tools you use. We’ll come back with a first live win, fixed scope and price.

Book a discovery call 30 minutes · no deck