Backend & APIs
Long-running services and the contracts between them.
Full-stack · Backend-leaning · Voice AI
Production systems where latency, scale, and correctness all matter at the same time.
Voice and real-time media pipelines, microservices that survive contact with the real world, and the AI plumbing that ties them together.
- Backend & APIs
- Voice AI & Telephony
- Real-time media
- Cloud & DevOps
Where there is depth
Expertise
A working toolkit, not a checklist. Each area has been shipped, debugged, and operated in production.
Voice AI & Telephony
End-to-end voice stacks — from microphone to model and back, including the clients that speak to them: an Alexa custom skill, a native watchOS app, an ESP32-S3 wake-word listener and a Chrome extension.
Real-time media
Moving audio and video through pipelines that have to keep up.
AI integration
Wiring LLMs into real products — observably, safely, on someone else's behalf. Also orchestrating specialised AI agents as a delivery team, with scoped write areas and evidence-based review gates before a change is accepted.
Modular AI systems
Swappable model providers behind one boundary: per-task routing (chat, embeddings, translation, speech, realtime), provider fallbacks and explicit "use current" defaults, credentials held in a single service, and every call traced with provider, model and token usage.
Cloud & DevOps
From container to cluster, with traces all the way down.
Data & persistence
Picking the right shape of storage for the problem at hand.
Selected work
Sample engagements
A few representative systems. Anonymized — what matters is the shape of the problem, not the logo.
01
Telephony stack for voice AI
Multi-stage Docker deployment of a Kamailio SIP proxy and RTPEngine media relay feeding a voice-AI agent. Tuned host networking and WSS termination for low-latency speech roundtrips. RBAC across multi-tenant APIs.
02
Self-hosted AI workspace
A central FastAPI backend serving two SPAs (a language-learning app and a markdown knowledge base), backed by Postgres with pgvector. A decoupled MCP server handles every LLM call — provider keys, request logs, OpenTelemetry traces and per-request event files all live behind one SSE boundary. The platform ships with a disposable Kubernetes test environment: chart smoke checks, an API suite and a voice suite driving real audio through the WebSocket pipeline, plus migrations proven against a restored production snapshot before release.
03
Healthcare microservices & CRM
CRM-facing services are extended, legacy data is migrated, code is reviewed (mostly backend, ~70–100% of the diff), and the architectural decisions shape the next quarter of work. Test coverage, CI/CD, Kubernetes via ArgoCD.
04
Video conferencing platform
Real-time video and audio using WebRTC, with backend services for voice recognition and translation. Infrastructure stood up with Terraform, CI/CD and Docker from a clean slate.
05
SaaS back-office for media clients
Tools for enterprise media customers, integrated with HubSpot, payments and e-invoicing providers. Lambda pipelines did the heavy media processing; Jenkins and Docker ran the show.
Contact
Let's talk
Backend, voice-AI and platform engagements. The fastest way to get in touch is email.