Orbit - Use Cases

🧩 Orbit: Use Cases & Reference Apps

Every AI pattern you need to know is built as a reference app on top of Orbit. Each one is a workflow definition + tools + prompts + an eval suite.

Coverage matrix

#	Reference app	Structured output	Tool calling	RAG	Workflow pattern	Agent	HITL	MCP	Multimodal	Trigger	Version
UC1	Support ticket triage	✅	✅		Routing					API	v1
UC2	Knowledge copilot (chat over docs)			✅	Augmented LLM + memory					Chat	v1 → v4
UC3	Refund agent with approvals	✅	✅	✅	Saga + compensation	✅	✅			API	v2
UC4	Research agent		✅	✅	Orchestrator–workers + evaluator–optimizer	✅		✅ web search		API	v4
UC5	Document extraction pipeline	✅			Map (fan-out) + validation retry		review queue		✅ PDFs/images	Cron + S3 upload	v4
UC6	Data analyst (text-to-SQL)	✅	✅ SQL tool	schema RAG	Plan → execute → verify	✅			charts	Chat	v4
UC7	SRE / DevOps agent		✅	runbook RAG	Routing + agent	✅	✅	✅ Prometheus/K8s		Alertmanager webhook	v4
UC8	Content pipeline	✅		✅ brand guide	Prompt chaining + gates + evaluator–optimizer		✅			API	v4
UC9	Meeting notes → actions	✅	✅		Chaining			✅ Slack/Jira	audio (stretch)	Webhook	v2
UC10	PR review bot	✅	✅	✅ code conventions	Parallelization (sectioning)	✅		✅ GitHub		GitHub webhook	v4
UC11	Multi-agent planner		✅	✅	Supervisor / handoffs	✅✅	✅	✅		API	v4
UC12	Voting classifier (safety/moderation)	✅			Parallelization (voting)				✅ images	Inline guard	v4

Use-case specs

UC1 · Support ticket triage (v1)

Input: raw ticket text/email → classify (category, priority, sentiment) with a JSON schema → extract (order ID, product, customer) → route: bug → create a GitHub issue (tool); billing → hand off to UC3; how-to → answer with UC2.

• Acceptance: ≥ 90% category accuracy on a 100-ticket labeled set; invalid JSON repaired within 1 retry; p95 < 3 s

UC2 · Knowledge copilot (v1 chat → v4 RAG)

Chat over a tenant’s docs, with citations, ACL-aware retrieval, conversation memory (summarized after N turns), and streaming.

• Acceptance: recall@5 ≥ 0.85 on 100 questions; faithfulness ≥ 0.9 (judge validated against human labels); says “I don’t know” when retrieval is empty

UC3 · Refund agent with approvals (v2)

Agent reads order history (tool) → checks policy (RAG) → proposes a refund → human approval (inbox, 24h timeout → escalate) → calls the payment tool → notifies the customer; on failure, compensates (reverse the ledger entry, reopen the ticket).

• Acceptance: kill workers mid-run 100× → no double refunds; every refund has an approver audit trail

UC4 · Research agent (v4)

Planner splits a question into 3–6 sub-questions → parallel sub-agents do web search (MCP) + reading → synthesizer writes a cited report → evaluator scores coverage/citations → revise ≤ 2 loops.

• Acceptance: token budget enforced; loop detection; report quality judged ≥ 4/5 on 20 prompts

UC5 · Document extraction pipeline (v4)

Upload invoices/resumes (PDF/image) → multimodal extraction → schema validation + business rules → low-confidence items go to a human review queue → store to Postgres. Runs a map step over 500 docs with bounded concurrency; nightly cron.

• Acceptance: field-level accuracy ≥ 95% on 50 labeled docs; resumable after a crash; cost per doc reported

UC6 · Data analyst, text-to-SQL (v4)

Question → retrieve relevant table schemas (RAG over the catalog) → generate SQL → read-only role + allowlisted tables + LIMIT + timeout → execute on ClickHouse → self-check the result → answer with a chart spec.

• Acceptance: 80% execution accuracy on 50 questions; zero writes possible (prove it with attack prompts)

UC7 · SRE agent (v4)

Alertmanager webhook → gather context via MCP (Prometheus queries, K8s events, recent deploys) → RAG over runbooks → hypothesis + proposed action → approval before any mutating action (e.g. rollout restart).

• Acceptance: correct root-cause category in 7/10 injected incidents from your chaos drills

UC8 · Content pipeline (v4)

Brief → outline → gate (checks structure) → draft → critique (evaluator) → revise → brand-style check (RAG over the style guide) → human approval → publish via webhook.

UC9 · Meeting notes → action items (v2)

Webhook with a transcript → summarize → extract action items (owner, due date) → create tasks via Jira/Slack tools (MCP) → post a summary. Idempotent on webhook retries.

UC10 · PR review bot (v4)

GitHub webhook → fetch the diff (GitHub MCP) → parallel reviewers (security, performance, style) → aggregate → post comments. Rate-limited per repo; conventions retrieved from the repo’s docs.

UC11 · Multi-agent planner (v4)

Supervisor agent delegates to researcher, coder (code-exec sandbox), and critic agents; shared state in LangGraph; human checkpoint before the final action. Compare quality and cost against a single agent.

UC12 · Voting classifier / guardrail (v4)

3 parallel cheap-model votes (or 1 cheap + 1 strong) for moderation / prompt-injection detection; used as an inline guard step by other workflows.

Platform capabilities checklist (what Orbit itself must support)

• Step types: llm, tool, http, code, rag.search, agent, branch, map, parallel, human, wait/timer, subworkflow
• Expression language for conditions/mappings ({{ steps.classify.output.priority == "high" }})
• Prompt registry with versions + A/B + eval scores
• Tool registry: JSON Schema inputs, auth, scopes, “requires approval” flag, timeouts, idempotency
• Triggers: API, cron, webhook, Kafka, S3 upload
• Run timeline + trace viewer (every LLM call, tool call, token count, cost)
• Budgets and quotas per tenant/workflow; model routing policies
• Evals: datasets, suites, judges, CI gate, online sampling
• Guardrails: input/output validators, PII redaction, injection detection, allow/deny lists