Question 1

How do you integrate AI into a carrier catalog and order flow via TMF Open APIs (300+ APIs) with Amdocs CES?

Accepted Answer

We treat TMF Open APIs as the integration contract — TMF620 (Product Catalog), TMF622 (Product Ordering), TMF629 (Customer), TMF641 (Service Ordering), and TMF666 (Account) are the bread and butter. Against Amdocs CES specifically we connect through the catalog-driven order-capture layer and the CES Eventing API for catalog change streams; the AI runtime publishes recommended-offer events to TMF688 (Event), and order-decomposition decisions land back through TMF640 (Service Activation). The model never edits the catalog or the order — it produces signed recommendations the order-management layer can accept, override, or escalate, which keeps you within Amdocs upgrade paths and TMF ODA conformance.

Question 2

How do you build AI analytics that are lawful-intercept-aware under CALEA and ETSI TS 102 232?

Accepted Answer

The intercept-related information (IRI) and content-of-communication (CC) paths are physically separated from the analytics plane. CALEA Section 103 mandates that the carrier — not a vendor — controls the intercept; we therefore never co-locate model training with the LI mediation function. Aggregated, anonymised features (cell-level KPIs, slice telemetry, CDR-derived counters) flow to the analytics enclave through a one-way diode-style egress with k-anonymity enforced upstream. ETSI TS 102 232 handover interfaces (HI1/HI2/HI3) stay inside the LEMF boundary. The result: AI models can score churn, fraud, and network anomalies on the same subscribers without ever touching IRI or CC, and the LI audit trail stays clean.

Question 3

When does an O-RAN use case belong in an xApp on Near-RT RIC versus an rApp on Non-RT RIC?

Accepted Answer

The decision is latency budget and data horizon. Near-RT RIC enforces a 10ms–1s control loop and operates on E2 service-model interfaces (E2SM-KPM for KPIs, E2SM-RC for radio control) — that is the home for traffic-steering, mobility-load-balancing, QoS optimisation per UE, and beam selection. Non-RT RIC sits in SMO with a >1s loop over the O1/A1 interfaces and is correct for AI/ML training, policy generation, energy-saving cell-sleep schedules, and any model that needs minutes-to-hours of context. We typically train on Non-RT, generate A1 policies, and deploy xApps as quantised inference modules conformant to O-RAN WG3 specifications; on tier-1 RAN platforms we have prototyped xApps in 6–10 weeks.

Question 4

How do you monitor 5G Standalone slice SLAs at the granularity carriers actually need?

Accepted Answer

5G SA SLA monitoring needs three planes correlated: user-plane (UPF counters, per-slice throughput and packet-loss from N3/N6), control-plane (AMF/SMF session events from the SBI), and RAN (per-slice PRB utilisation, BLER, MCS distribution from gNB OAM). We stream these through a per-slice feature store keyed by S-NSSAI (SST + SD), compute sub-second percentile SLOs (typical: P99 latency under the contracted bound), and emit ETSI ZSM closed-loop intents when a slice drifts. The same telemetry feeds capacity forecasting so the network team sees a slice SLA breach forecasted hours before it would happen, not after.

Question 5

How does your IRSF and Wangiri fraud detection actually work at carrier volumes?

Accepted Answer

IRSF (International Revenue Share Fraud) and Wangiri ("one-ring") share a signature: short-duration calls to specific high-tariff destinations or premium ranges, often correlated across many subscribers in a narrow window. We stream CDRs (and increasingly IPDRs from VoLTE/VoNR) into a feature pipeline that maintains rolling counters per A-number, B-number prefix, originating cell, and IMSI — typical state size is hundreds of millions of keys. A gradient-boosted scorer flags candidate fraud in sub-second; a graph layer correlates SIM, device IMEI, and payment instrument to surface the operator ring. Carrier benchmark is to keep IRSF + Wangiri loss under 1% of voice revenue — production deployments have driven the loss curve below that band by combining the scorer with automatic premium-range hot-listing and a routing-layer block at the IGW.

Question 6

How do you do churn prediction and retention without violating EU EECC subscriber-rights provisions?

Accepted Answer

EECC Articles 105–107 cap contract length, force pro-rata exit, and require neutral comparison information — and Article 102 transparency rules mean retention scripts cannot be "dark-pattern" outreach. We build the churn model on usage and care-interaction features that are operationally necessary (network quality the subscriber actually experienced, ticket history, plan-fit), keep purpose-limited consent records per Article 6 GDPR, and route every retention offer through a fair-treatment policy layer: no offer materially worse than the subscriber would get walking back in cold, no offer hidden from a comparison portal. The agent script is grounded against the live catalog through TMF620, so the retention agent literally cannot offer a non-existent or non-portable bundle.

Question 7

Does cell-sleep network energy optimisation actually save money — and how do you wire it into TMF + ETSI ZSM?

Accepted Answer

On dense urban RANs we have seen RAN OPEX energy reductions in the 15–25% band by combining symbol-level micro-sleep, channel-level deactivation, and carrier-level shutdown during demonstrated low-traffic windows — and energy is typically 20–40% of RAN OPEX, so the math compounds. The ML side forecasts per-cell traffic 15–60 minutes out using a Non-RT RIC rApp; the orchestration side emits A1 policies into the RAN and ETSI ZSM closed-loop intents into the cross-domain orchestrator, with a TMF640 service-activation reversal path if any slice SLA degrades. Guardrails are non-negotiable: no SLA-bearing slice ever gets traded for energy savings, and the policy auto-reverts on KPI drift.

Question 8

Can you do document AI on interconnection agreements, LOAs, and wholesale contracts at carrier scale?

Accepted Answer

Yes — this is one of the highest-ROI document AI workloads in telecom because the contracts are dense, custom, and the obligations (settlement rates, MEF 3.0 LSO commitments, lawful-intercept clauses, SLA penalty schedules) drive real money. We extract structured obligations with a layout-aware NER pipeline tuned for telecom legalese (interconnect points, CIC codes, E.164 ranges, MEF service IDs, hand-off latency commitments), reconcile against the live wholesale catalog, and surface drift between contracted terms and what the OSS/BSS is actually billing or provisioning. Letters of Authorization (LOAs) for number porting and circuit migration go through the same pipeline and feed the porting workflow with extracted entities. Typical pilot: 4–6 weeks on a sample of 200 agreements, full deployment in 8–12 weeks.

Carrier-grade AI: TMF + ODA-native, O-RAN xApps, IRSF detection, 5G SA slice SLAs

Frameworks we build to

Carrier-grade AI Systems

O-RAN xApps & rApps

IRSF + Wangiri + SIM-swap Detection

Churn + EECC-compliant Retention

5G SA Slice SLA Monitoring

Cell-sleep Energy Optimisation

Field-service Dispatch + Trouble-ticket AI

Document AI: Interconnect + LOAs

TMF-grounded Care Agent-assist

Telecom Data Engineering

CDR / IPDR Streaming Pipelines

Per-slice Telemetry Lake

Subscriber 360 over TMF Customer Domain

Billing & Mediation Analytics

M2M / IoT eUICC Device Plane

Lawful-intercept-safe Egress

Frequently Asked Questions

How do you integrate AI into a carrier catalog and order flow via TMF Open APIs (300+ APIs) with Amdocs CES?

How do you build AI analytics that are lawful-intercept-aware under CALEA and ETSI TS 102 232?

When does an O-RAN use case belong in an xApp on Near-RT RIC versus an rApp on Non-RT RIC?

How do you monitor 5G Standalone slice SLAs at the granularity carriers actually need?

How does your IRSF and Wangiri fraud detection work at carrier volumes?

How do you do churn prediction and retention without violating EU EECC subscriber-rights provisions?

Does cell-sleep network energy optimisation actually save money — and how do you wire it into TMF + ETSI ZSM?

Can you do document AI on interconnection agreements, LOAs and wholesale contracts at carrier scale?

How long does each of these typically take to ship?

Carrier-grade AI, shipped against your BSS / OSS and RIC