Solutions
Four solution tracks organized by business problem
The same adjudication capability, delivered in combinations matching each customer’s data boundary, disposition authority, and business workflow. Each track states the problem, approach, workflow, and current scope.
Carrier anti-fraud
Business problem
Port governance and user outreach are statutory duties, but rule engines keep failing against variant scripts; running everything through a large model is unaffordable in cost and latency; blocking actions demand extreme precision — blocking an official message triggers complaints immediately.
Approach
The SMS anti-fraud gateway: signal, then content, then user — three decision layers; 17 regex classes plus nine parallel questions at the content layer; 90%+ of traffic handled by rules/reputation with the model only on the 5–10% candidate set; zero disturbance to the 12381/96110 whitelist.
Deliverables
- Port profiles & governance lists
- Four-level dispositions
- User alert event stream
- 12381 / 96110 coordination
The SMS gateway is demo-environment scope (46,200/day, 82ms, zero official false alerts); IM intelligence is a real batch-run capability. Delivery follows the shadow-then-inline staged path.
Telecom risk analytics
Business problem
IM, SMS, web, and app systems each fight alone: entities do not interconnect, dispositions do not close the loop; the same fraud family leaves the same entities across channels, but single-point systems never see the whole.
Approach
A cross-scenario entity graph plus one adjudication engine: normalize domains, cards, accounts, apps, and sites into entities; correlate families and expand leads on the graph; the decision model provides the unified fraud-judgment layer.
Deliverables
- Family graph & expansion leads
- Black/gray group profiles
- IOC entity lists
- Evidence packs & forensic reports
TG group intelligence is a real batch run (58,133 groups, 99.1% agreement); web/app analysis capability ready; the cross-scenario platform is partially delivered.
Intelligence & forensics
Business problem
Analysts dig through multiple systems: scattered leads, hand-assembled evidence chains, conclusions that are hard to trace; cross-border and cross-language intelligence is expensive to organize.
Approach
The analyst Copilot direction: GraphRAG treats entities, relations, and evidence as retrieval objects; natural language becomes structured queries; output is conclusion + evidence chain + confidence, exportable as a forensic pack. Cross-language capability builds on the model base’s 201-language coverage.
Deliverables
- Natural-language research queries
- Evidence chains & confidence
- Exportable forensic packs
- Cross-language lead normalization
The Copilot is a planned direction; the underlying entity graph and NL-to-query capabilities exist; the product is in build. The intelligence API is a planned lightweight form.
Device-side protection
Business problem
The manipulation stage happens on the victim’s phone: call steering, screen sharing, fraud apps. Cloud solutions cannot see the content, and post-hoc blocking is too late.
Approach
On-device real-time alerts: line-by-line call adjudication (<50ms demo metric), five-dimensional fraud signals, behavioral sequence alerts (screen sharing / silence / payment launch); content never leaves the device — only risk conclusions; family linkage and insurance add-ons.
Deliverables
- Real-time call risk alerts
- Behavioral sequence alerts
- Family notifications
- Anti-fraud insurance add-ons
Training data assets ready (28,511 voice-text pairs, 12,406 elder dialogs); on-device <50ms is a demo metric; productization in progress.
Business model
Four revenue lines, from projects to subscriptions
Delivery forms combine along the customer’s data boundary, disposition authority, and payment habits.
Carrier provincial projects
ProjectSMS gateway / anti-fraud platform build, expansion, and operations; coordination with 12381/96110 and roaming alerts.
Bank / payment money-flow
Project + usageFraud account, QR code, and fund-path monitoring + risk subscription API.
Enterprise SMS compliance API
Usage / annualSignature, template, link, and script self-checks.
Elder protection / insurance
Subscription / policy shareOn-device alerts + family linkage + insurance add-ons.
Business model partially validated; specific contracts and customer cases pending formal disclosure.
Return on investment
Not “one more system” — “one less loss”
Three levers: compute, labor, and loss; example calculations to be replaced with customer figures.
Compute savings
90%+ of traffic handled by rules/reputation; the model sees only the 5–10% candidate set — an order of magnitude below full-volume LLM inference.
Labor savings
Bimodal probability distribution keeps the ambiguous band at ~6%; review volume stays bounded; 99.1% run-to-run agreement reduces disputes.
Loss reduction
High-risk users reached earlier; zero disturbance on official whitelists; appealable and reversible actions control secondary risk.
Calculations are example scope (Data class: scenario) — replace with customer-specific figures.
Start with one track
Pick the track closest to your business boundary; complete offline validation in 2–4 weeks, then decide the scaling path.