Cold Chain Sentinel: AI Voice Triage

Autonomous cold-chain voice triage engine intercepting reefer spikes via CALL-E driver calls, deterministic 8-node LangGraph routing, and Gemini 3.5 Flash cognitive audits.

Comment

Inspiration

In commercial refrigerated logistics, a single mechanical failure—such as Reefer Compressor Alarm 18 (High Engine Temp)—can spike trailer internal temperatures from a -20.0°F frozen setpoint to -6.2°F while hauling $180,000 of temperature-sensitive biologics and vaccines. Fleet dispatchers face a critical 25-minute allowable excursion window before catastrophic cargo spoilage occurs.

Traditional fleet management systems fail during this window:

  1. Standard telematics alerts sit unread in crowded email inboxes for 15 to 45 minutes.
  2. Fatigued dispatchers play chaotic phone tag with drivers, misquoting alarm diagnostic codes.
  3. Rerouting drivers without verifying their Electronic Logging Device (ELD) status risks an FMCSA 49 CFR Part 395 civil penalty of up to $16,000 per violation.
  4. FDA Food Safety Modernization Act (FSMA 21 CFR Part 1, Subpart O) mandates tamper-proof, timestamped temperature-control audit trails that email threads cannot provide.

We built Cold Chain Sentinel to replace passive telemetry dashboards with an active, autonomous voice triage engine. By fusing the CALL-E telephony API with an 8-node deterministic LangGraph state machine and Google Gemini 3.5 Flash, Sentinel directly calls commercial drivers on the highway within 4 seconds of a thermal spike, audits the situation, and safely actuates an emergency warehouse reroute.

What It Does

Cold Chain Sentinel acts as an autonomous tier-1 dispatch specialist for cold-chain fleet operations:

  • Sub-4-Second Autonomous Ingress: Intercepts IoT telematics excursion webhooks, validates E.164 phone formats, and enriches load manifests with TMS bills of lading and ELD Hours of Service (HOS) in parallel.
  • Autonomous In-Cab Driver Interrogation: Dispatches an automated phone call via CALL-E directly to the driver's cab. Using a structured IVR checklist, it verifies physical alarms (Alarm 18, compressor head pressure, evaporator airflow) and confirms remaining legal drive time.
  • Grounded Cognitive Compliance (Gemini 3.5 Flash): Ingests the verbal triage transcript at temperature=0.0, strictly structured via Pydantic V2 schemas (ComplianceReviewDecision). Every risk score cites explicit state evidence (e.g., cargo sweating, bulkhead blockage), backed by dynamic fallback to Gemini Pro and prompt-injection defense (OWASP LLM01).
  • Deterministic Fleet Actuation: If all safety gates pass, Sentinel executes three atomic actuation legs in under 4 seconds: mutating the fleet GPS route to the nearest cold hub (e.g., Lincoln Cold Logistics Hub), reserving an emergency dock bay, and sending an SMS dispatch confirmation to the driver.
  • Fail-Safe P0 Human Escalation: If international cellular drops (SIP 408), VoIP timeouts, driver refusals, or unverified diagnostics occur, Sentinel deterministically halts autonomous actuation, locks a mathematical monotonic-or safety latch, and alerts human dispatchers with a high-contrast P0 emergency docket.
  • Task-First Operations Cockpit: A Next.js 16 dark glassmorphism dashboard streaming 9 typed Server-Sent Events (SSE) in real-time, rendering generative timeline cards and telemetry gauges.

How We Built It

Cold Chain Sentinel is architected with a strict dual-track submission boundary:

1. The Core 8-Node LangGraph State Machine (src/agents/)

We implemented an acyclic LangGraph StateGraph that enforces strict separation of concerns across 8 sequential nodes: ingress ➔ enrichment ➔ call_interrogation ➔ compliance_review ➔ decision_gate ➔ [autonomous_actuation | escalation_failure] ➔ persistence_audit. Crucially, cognitive reasoning nodes hold zero fleet tool bindings (OWASP LLM06 Excessive Agency prevention). Routing decisions are pure Python conditional edges, never non-deterministic LLM recommendations.

2. State Safety Invariants & Pure Reducers (src/state/reducers.py)

All state is managed via SentinelState with custom pure reducer channels:

  • monotonic-or Reducer: Governs requires_immediate_human_override. Once any node, telephony failure, or safety gate flips this flag to True, it is mathematically impossible for any subsequent tool, node, or LLM to reset it to False.
  • immutable-after-init: Protects entry telematics and cargo manifests from tampering.
  • merge-by-key & append-only: Ensures concurrent tool artifacts and regulatory audit events never clobber each other.

3. Telephony Integration & Resilience (src/tools/call_e_client.py)

We integrated the CALL-E SDK using client.calls.create_and_wait wrapped inside asyncio.to_thread for non-blocking FastAPI async execution. We engineered:

  • Zero-Redial Telephony Invariant: A strict pre-call guard checking call_id is None. If a call drops or fails, Sentinel immediately routes to human escalation rather than looping, conserving live credits and preventing driver harassment.
  • Dynamic Telecom Sanitization: E.164 phone validation, schema normalization (stripping $defs and unsupported unions), and dynamic regional routing (IN for +91, US default).

4. Operations Cockpit Frontend (frontend/)

Built with Next.js 16, React 19, and Tailwind CSS v4, consuming an SSE endpoint (/sessions/{event_id}/stream) with 9 typed event models. Features real-time generative UI cards: EnrichmentSummaryCard, InterrogationResultCard, RiskAssessmentCard, ActuationTransactionCard, and EscalationAlertBanner.

5. Air-Gapped CALL-E Agent Skill (skills/cold-chain-reefer-triage/)

In addition to the orchestrator, we authored a self-contained, zero-framework, dependency-free agent skill (triage.py + SKILL.md) designed for direct upstream PR integration into CALLE-AI/awesome-phone-call-agents.

Challenges We Ran Into

  1. International Telecom Carrier Drops (SIP 408): When dispatching live calls from our development environment to Indian phone numbers (+91), local telecom anti-spoofing ILD firewalls dropped unregistered automated VoIP packets with SIP 408 / NO ANSWER. Instead of treating this as a blocker, we turned it into an enterprise feature: we built a deterministic timeout catcher that triggers an automated, fail-safe P0 Human Escalation, and validated the happy path using CALL-E's official US test hotline (+1 276-322-9632).
  2. Preventing LLM Hallucinations in Safety Decisions: LLMs should never have direct execution authority over physical trucks. We eliminated hallucinations by stripping all tool bindings from Gemini 3.5 Flash and restricting its role to structured citation extraction (ComplianceReviewDecision). Actuation tools are only reachable after our deterministic Python decision_gate evaluates six strict criteria (FMCSA HOS limits, cargo sweating, telephony completion, driver refusal, injection checks).
  3. Checkpointer State Resumption Without Redialing: In a server crash mid-incident, LangGraph state recovery must not re-dial the driver. By checking transaction logs in SQLite and PostgreSQL (SqliteSaver / PostgresSaver), Sentinel resumes cleanly from the exact completed node without initiating duplicate calls.

Accomplishments That We're Proud Of

  • 88 / 88 Passing Automated Regression Tests: 100% test pass rate across 12 suites in 5.20s—covering graph topology, monotonic reducers, OWASP prompt injection, telephony fallbacks, and full end-to-end webhook pipelines.
  • Verified Live Telephony Execution: Documented in our live CALL-E billing ledger (dashboard.heycall-e.com), proving a real 1-minute 24-second autonomous call (Call ID: 23db451665034286abd8c30a3414a56d) over production trunks.
  • Sub-4-Second End-to-End Fleet Reroute: Ingesting telematics, conducting driver triage, performing Gemini risk audits, and committing cross-dock reservations in under 4 seconds post-call.
  • Regulatory Alignment: Built from day one to comply with FDA FSMA 21 CFR Part 1 audit requirements and FMCSA 49 CFR Part 395 HOS driver safety mandates.

What We Learned

  • Determinism Over Generative Chaos: In high-liability domains like cold-chain pharma, deterministic state graphs with mathematical reducers (monotonic-or) are far superior to open-ended ReAct agent loops.
  • Voice is the Ultimate Bridge: Commercial drivers cannot navigate complex apps while driving at 65 mph in freezing rain. Natural, structured voice phone calls via CALL-E bridge the physical highway and digital dispatch systems seamlessly.
  • Defense-in-Depth for Telephony: Production voice AI requires immediate fail-safe escalation protocols the moment VoIP packets drop or carrier timeouts occur.

What's Next for Cold Chain Sentinel

  • Multi-Carrier Telephony Failover: Implementing automated secondary carrier trunking if primary cellular routes drop with SIP 408 errors.
  • Direct TMS/ELD Enterprise Connectors: Native certified webhooks for Samsara, Motive, Geotab, and McLeod Fleet Telematics.
  • Multi-Lingual In-Cab Triage: Expanding CALL-E real-time voice prompts to Spanish, Punjabi, and French Canadian dialects for diverse North American freight corridors.
  • Edge Deployment: Packaging Sentinel's 8-node state machine to run on edge compute gateways inside modern smart reefer trailers.

Built With

  • asyncio
  • call-e
  • fastapi
  • google-gemini
  • iot
  • langgraph
  • llm
  • logistics
  • next.js
  • owasp
  • postgresql
  • pydantic
  • pytest
  • python
  • react
  • rest-api
  • server-sent-events
  • sqlite
  • tailwind-css
  • telematics
  • typescript
  • uvicorn
  • voice-ai
  • webhooks
Share this project:

Updates

Submission history