The agents we ship most often.

Responsibilities

• Multi-channel engagement (web chat, email reply, SMS)
• Structured qualification (pain, timeline, authority, budget signal)
• Calendar booking for qualified leads
• Nurture routing for not-yet-ready leads

Building blocks

• LangGraph for branching qualification flow
• Calendar integration (Cal.com / Calendly)
• CRM write-back (HubSpot / Salesforce)
• Eval suite for qualification quality

Guardrails

• Never quote pricing — always defer to the AE
• Escalate to a human SDR if the lead is angry or expresses urgency
• Log every conversation with intent classification

Responsibilities

• Answer documented questions with citations
• Triage and classify inbound for engineering escalation
• Capture diagnostic context (logs, screenshots, network traces)
• Hand off to humans with full conversation context

Building blocks

• Hybrid search (BM25 + vector)
• Reranking (Cohere Rerank v3)
• Citation-required prompting
• Eval suite for hallucination + refusal correctness

Guardrails

• Refuse to make policy promises — defer to humans
• Refuse low-confidence answers and route to a human
• Surface contradictions between sources rather than picking one

Responsibilities

• Watch production / maintenance / quality / inventory / energy
• Generate shift handoff briefings
• Route alerts to the right human via WhatsApp + dashboard
• Coordinate across agents on cross-cutting incidents

Building blocks

• LangGraph orchestration per agent
• Shared Postgres event bus
• Kafka stream from SCADA / IoT / ERP
• Edge gateway for on-prem reasoning

Guardrails

• No agent takes physical action — humans always pull the trigger
• Every alert tied to source data with audit trail
• Confidence thresholds tune per agent and machine class

Responsibilities

• OCR + vision extraction from PDFs and scanned docs
• Validate extracted fields against business rules
• Route confidence-below-threshold to a human reviewer
• Post clean records to downstream systems

Building blocks

• Vision LLM for layout-rich docs
• Structured-output prompting with explicit confidence scores
• Business-rule validation layer
• Eval suite with held-out historical docs

Guardrails

• Never normalize person names without explicit approval
• Confidence < 0.7 means human review, no exceptions
• Preserve raw extracted strings for audit

Responsibilities

• Greet and qualify in under 30 seconds
• Answer common questions in 1-2 short turns
• Route to humans on emotional signals or complexity
• Log every call with summary + sentiment

Building blocks

• Vapi / Retell for low-latency streaming
• Twilio for telephony
• RAG over knowledge base + caller history
• Real-time sentiment detection

Guardrails

• Never pretend to be human
• Never make commitments about pricing, refunds, or SLAs
• If silence > 5s, prompt 'still there?'

Responsibilities

• Translate plain-English questions into ERP API calls
• Background anomaly detection (cost drift, stock ageing)
• Suggestion engine with approve / reject workflow
• Audit trail tying every response to source records

Building blocks

• LangGraph with explicit tool definitions per ERP module
• MCP server architecture for tool composability
• Prompt caching for the schema context
• Approval workflow for cost suggestions

Guardrails

• Never auto-post to the ledger — always route to a named approver
• Cite source records on every response
• Refuse cross-tenant questions in multi-tenant deployments

Responsibilities

• Extract structured fields from KYC documents
• Sanctions and PEP screening across global lists
• Risk scoring with model card + reasoning trace
• Route exceptions to compliance analysts

Building blocks

• Vision LLM for ID + document parsing
• Sanctions API integration
• LightGBM for risk scoring
• Approval workflow with named approver chain

Guardrails

• Never auto-approve high-risk cases — always require human review
• Log every decision with model version + reasoning + approver
• Encrypt PII in transit and at rest, retention per jurisdiction

Responsibilities

• Real-time defect detection on production lines
• Active learning loop — uncertain cases to humans
• Aggregate defect-pattern reporting
• Integration with line-stop authority signals

Building blocks

• YOLOv10 for fast common-defect detection
• Vision LLM for novel / ambiguous defects
• Active-learning sampling for the uncertain bucket
• Edge gateway for low-latency inference

Guardrails

• Never auto-stop the line without supervisor confirmation
• Surface confidence per detection
• Retrain weekly with the active-learning sample

Responsibilities

• Pull access reviews, change tickets, training records from source systems
• Format per auditor's template
• Upload to evidence platform (Vanta / Drata) or Drive
• Notify the GRC lead on anomalies

Building blocks

• n8n orchestration
• API integrations to identity, ticketing, HR, training systems
• Structured-output prompting for normalization
• Eval suite that replays past audit cycles

Guardrails

• Never modify source-of-truth records
• Always preserve raw evidence alongside formatted output
• Surface gaps in evidence collection rather than fabricating