Telecom and commercial risk treated as first-class engineering
The hardest parts were the local caller ID, the multi-vendor call economics, and the carrier-billing clause that decides the margin. I handled them with the same rigour as the software.
Case study · Flagship engagement
Samleng: AI voice notifications for enterprise logistics.
How an automated native-language calling platform was designed, built and load-tested to a contract-ready state in about 7 weeks, and why the client paused before rollout, so it never ran in production. The flagship engagement behind Motta Thong's AI systems work.
Public, de-identified version. Client de-identified; no pricing, margins, or stack internals. Figures are engagement design figures, validated pre-rollout. The full named technical version is available under NDA.
At a glance
- Client
- One of Southeast Asia's largest logistics and parcel networks (named under NDA)
- Sector
- Logistics / last-mile parcel delivery
- Challenge
- A manual, branch-based operation calling customers by hand to announce parcel arrivals
- Scope
- Discovery → solution design → telecom procurement → financial modelling → cloud architecture → build & load test → rollout plan
- Delivery model
- AI-native: a lean senior team directing AI agent fleets under human architecture and review
- Modelled outcome (not measured)
- The financial model projected a reduction in notification cost. The platform never ran in production, so this was never realised and no saving was banked
- Engineering proof
- 181/181 automated tests with zero dependency vulnerabilities; outbound dialling layer load-tested to 20,000 concurrent calls
- Timeline
- ~7 weeks from first engagement to a built, validated platform and a contract-ready rollout
- Status
- Paused before contract. The client did not sign and the platform was never put into production
Summary
A market-leading logistics network was notifying customers of parcel arrivals entirely by hand, branch by branch, at a large recurring cost, and an internal attempt to automate it had stalled.
Samleng places parcel-arrival calls in natural, native-language voice, triggered automatically the moment a parcel is scanned at a branch. The financial model, stress-tested against deliberately unfavourable assumptions, projected a cut in the cost of the operation. The engagement was full-scope: opportunity discovery and business-case construction, a solution scoped against a complete call-flow audit, live telecom procurement to secure a local caller ID, a defensible cost-and-margin model, a documented enterprise cloud architecture, and a production-grade governance package.
What did not happen: the client paused before contract. They never signed, and the platform never carried production traffic. Every cost figure here is a modelled projection against their own baseline, not a measured result. No saving was banked, no roles changed, and none is claimed. What the engagement genuinely produced was a decision-ready view of the economics, the vendor risks, the security gates and a safe launch sequence.
7 phases to contract-ready · rollout on signatureWeeks
Phase
123456789101112
1 · Discovery
2 · Solution & MVP
3 · Telecom procurement
4 · Financial model
5 · Cloud architecture
6 · Deploy & load test
7 · Rollout & governance
Delivered to contract-ready (Weeks 1 to 7)Begins on contract signature
The approach
- 01
Engagement and discovery
Rebuilt the cost baseline from the ground up, branch count, staffing, per-agent cost, and a defensible regional loading multiplier, so the business case would survive a CFO's scrutiny. Mapped how the decision would actually be made and where the real competition was: the client's own stalled internal build.
- 02
Solution scoping and MVP
Scoped an event-driven product: a parcel is scanned, an event reaches the platform, a native-language voice message is rendered and the call is placed within seconds. Scoping included a full end-to-end call-flow audit, a multi-agent review that surfaced 8 critical, 13 high, and 12 medium failure modes, each given a chosen fix before any real customer was dialled. A separate red-team pass triaged 61 issues into a hardening backlog.
- 03
Telecom procurement
Placing automated calls with a genuine local caller ID is a telecom problem, not a coding one. Ran a live procurement across the call-platform layer and the local carrier market, took a competitive carrier offer to a signed quotation, and designed a multi-vendor path so no single provider owns the whole stack, surfacing the single commercial variable that governs the platform's economics and flagging it for written resolution.
- 04
Cost and financial model
Built a model designed to survive a CFO and to tell the truth about the economics, running multiple cost scenarios rather than one flattering case. It projected a saving and isolated the one variable that could undermine it, early, while it was still cheap to address. A projection, not a banked result: the platform never reached production.
- 05
Cloud architecture
Produced an enterprise target architecture documented to an auditable standard: 10 formal Architecture Decision Records; service-level objectives engineered above the published SLA, with a 1-hour RTO and a 5-minute RPO. Validated with the platform vendors' own solution engineers, then pressure-tested through structured review and an independent automated audit.
- 06
Deployment and load test
The platform passes 181 of 181 automated tests with zero dependency vulnerabilities. The outbound dialling layer was soak-tested to 20,000 concurrent calls, roughly 250× the launch ceiling, at a 99th-percentile latency of 76 milliseconds, with no errors. A reusable load suite was built for every future scale step.
- 07
Production rollout and governance
The enterprise wrapper is complete and contract-ready: a phased, gated migration plan with explicit go/no-go criteria and a parallel-run period; a 25-item risk register with named owners and mitigations; and a governance layer including a client-facing data-security overview and a pilot security checklist.
Call-flow audit: failure modes found before launch+ 61 red-team issues triaged
P0 · Critical
8
Each received a chosen remediation
P1 · High
13
Prioritised in the hardening backlog
P2 · Medium
12
+ 6 false positives, dismissed
Event-driven call path: scan to dialled callMulti-vendor architecture
BranchParcel scanEvent fires on arrival
SamlengSigned webhookHMAC-SHA256 · idempotent
VoiceNative renderValidated · cached by hash
PlatformCall controlProgrammable dial layer
CarrierLocal originationBYOC trunk · local caller ID
Built or orchestrated by meVendor service
Engineering proofBuilt & load-tested
20,000Concurrent calls · ≈250× cap
181/181Automated tests passing
76 msp99 call-creation latency
0Dependency vulnerabilities
Risk register: severity-rated, named ownersExcerpt · 7 of 25
| ID | Risk & mitigation | Severity | Owner | Status |
|---|---|---|---|---|
| R-01 | Carrier bills per call attempt, not on connect Confirm billing basis in writing before any full-volume commit; secure wholesale rates. A roughly 4× swing in cost of goods, the single variable that governs the platform's economics. | Critical | Commercial | Open · pre-signature |
| R-02 | Answering-machine detection races audio playback Deferred-playback design: a real person never hears the message start late. | Critical | Voice eng | Resolved |
| R-03 | No local caller ID from a single vendor Bring-your-own-carrier path over a signed local SIP-trunk quotation secures origination and a genuine local caller ID. | High | Telecom | Mitigated |
| R-04 | Platform capacity at full daily volume unconfirmed Vendor solution-engineer confirmation pending in writing; outbound layer soak-tested to 20,000 concurrent. | High | Telecom | Open |
| R-05 | DB-pool / webhook-handler ceiling unmeasured Throwaway database load test scoped and sequenced; not yet run. | High | Platform eng | Open |
| R-06 | Webhook double-counts a call on carrier retry HMAC-SHA256 signature plus a deduplication key; ingestion is idempotent. | High | Platform eng | Resolved |
| R-07 | Client-side signature & rollout delay Phased paid ramp (first branches) keeps the engagement profitable through any delay. | High | Engagement | Monitored |
OpenResolved / mitigatedMonitored
Showing 7 of 25 · illustrative excerptResults
- A modelled reduction in the cost of customer notification, stress-tested against unfavourable assumptions. Modelled, never realised: the platform did not go to production.
- A built, tested platform: 181/181 automated tests passing, zero dependency vulnerabilities, and an outbound layer proven to 20,000 concurrent calls.
- A working local-caller-ID telephony path secured through real carrier procurement and a signed quotation, solving a problem off-the-shelf providers could not.
- A decision-grade financial model that isolated the one variable governing profitability, carrier billing basis, before it could be discovered the expensive way.
- An audit-ready architecture package, 10 ADRs, SLOs and a 25-item risk register, validated with the platform vendors' solution engineers.
- Reached a built-and-validated platform with a staged rollout in about 7 weeks, after an internal build had stalled. The client then paused before contract, so it never went live.
Why it worked
Modelled to find the truth, not to win the room
The model surfaced the scenario that could lose money and named it the headline risk. Enterprise buyers trust the person who shows them the downside.
Documented to a standard buyers can audit
Decision records, explicit SLOs, a named-owner risk register, and an NDA-gated technical review: what serious procurement asks for and rarely gets from a boutique.
AI-native delivery is a force multiplier
A lean senior team orchestrating AI agent fleets, for build, adversarial review, research and independent audit under human direction, reached enterprise depth across software, telecom, finance, and governance.
Capabilities demonstrated
- Opportunity discovery and business-case construction that survives CFO scrutiny
- Telecom and vendor procurement, including carrier and SIP negotiation in emerging markets
- Financial and margin modelling that isolates the variables that matter
- Audit-ready cloud solution architecture (decision records, SLOs, risk registers)
- Production engineering and load testing of real-time, high-concurrency voice systems
- Enterprise security and governance packaging
Join the community
I help people become the one their company trusts to run AI work. AI for Consultants is my free community on Skool.
Join AI for Consultants free (opens skool.com in a new tab)
Doors open 30 September 2026. Join free today and you are in on day 1.
Prepared with Ongkrong Consulting. The client paused before contract and the platform was never put into production, so every cost figure on this page is a modelled projection and no live saving, cost reduction or headcount outcome is claimed. Client details available to under NDA. Accurate as of 19 September 2026. Read the engagement write-up on ongkrong.com.