Selected work
Production systems, not portfolio mockups.
Client names stay confidential — that's the nature of infrastructure work for banks and government agencies. What we can show is the shape of the systems and the load they carry.
National-scale contact center platform
A high-throughput contact center stack serving hundreds of concurrent agent seats. We designed the architecture, built the telephony core, and operate the platform around the clock.
- Scope: architecture, FreeSWITCH core build, Kamailio edge routing, agent portal, 24/7 managed operations
- Load: 500+ agent seats, thousands of concurrent sessions at peak
- Constraint: carrier-grade availability with engineered failover across sites
- Outcome: platform in continuous production operation, run by the team that built it
High-volume messaging gateway
SMPP-level SMS infrastructure carrying millions of messages per day — OTPs, alerts, and two-way customer traffic — with route quality engineered at the protocol level.
- Scope: gateway build, carrier interconnection, throttling and queueing, DLR reconciliation
- Load: 10M+ transactions daily across interconnected routes
- Constraint: burst tolerance for campaign and alert storms without queue collapse
- Outcome: measurable route quality — delivery reporting the client can audit per route
Voice AI agents on production telephony
LLM-driven voice agents placed directly inside a live FreeSWITCH call path — streaming speech recognition, real-time reasoning, and natural TTS, inside a strict latency budget.
- Scope: media-path integration, streaming STT (Deepgram), LLM orchestration, Cartesia TTS, human handoff
- Load: live production call traffic, not a sandbox demo
- Constraint: sub-300ms response budget end to end; deterministic fallback on model failure
- Outcome: agents handling real calls with warm transfer to humans when confidence drops
02 — Notes from the field
What we write down so you don't have to learn it the hard way.
Sub-300ms voice AI: where the latency actually goes
LatencyA budget breakdown across STT, LLM, and TTS legs — and which two optimizations matter more than everything else combined.
Failover runbooks that survive 03:00
RunbooksThe difference between a runbook that looks good in a wiki and one an engineer can execute half-asleep during a trunk outage.
Data-sovereign voice: architecture for in-country deployment
SovereigntyHow to keep call media, recordings, and transcripts inside jurisdiction without giving up redundancy.
SMPP throughput is a queueing problem, not a bandwidth problem
SMPPWhy gateways fall over during OTP storms, and the backpressure design that prevents it.
Your system could be the next one in production.
Tell us the load, the constraints, and the deadline. We'll respond with a concrete assessment.