From signal path to recovery path.
- WebRTC, LiveKit, audio/VAD/STT/LLM/TTS pipelines
- Typed tools, queues, cancellation, and degraded modes
- Memory retrieval, provenance, policy, and compaction
- OpenTelemetry, traces, metrics, and operator surfaces
Lead software engineer · AI systems
Mert Sefa Akgün
I build the systems that make real-time AI feel immediate, interruptible, and trustworthy.
I design and ship real-time AI agents, voice systems, memory infrastructure, and the observable runtimes underneath them.
Real-timelatency and interruption ownership
Agenticbounded tools and resumable state
Observablestage telemetry and failure evidence
Systemsembedded roots to production runtime
Selected systems
Claims backed by architecture.
Three proof lanes show the current specialization, the architecture thesis behind it, and the lower-level systems roots underneath both.
production 01
NDA-safe architecture from production work
Real-time voice agent runtime
A production-oriented LiveKit pipeline spanning audio ingress, VAD, STT, LLM/tool mediation, TTS, interruption recovery, memory side paths, and stage-level telemetry.
proof surfaceLatency ownership · bounded queues · degraded modes · OpenTelemetry
Inspect artifact →implemented 02
Personal agent infrastructure
Provenance-first agent memory
A typed memory policy where source, trust, sensitivity, allowed use, expiry, and auditability stay attached to retrieval instead of disappearing into embeddings.
proof surfacePolicy evaluator · conservative defaults · side-effect gates · tests
Inspect artifact →field archive 03
Embedded/RF systems roots
Indoor localization with UWB
A field-system design combining UWB ranging, Bluetooth provisioning, and GPS/time sync under access-control hardware constraints.
proof surfaceRadio boundaries · calibration · deployment constraints · hardware notes
Inspect artifact →Field notes & artifacts
Writing as technical evidence.
Architecture notes expose constraints, owner boundaries, failure modes, and trade-offs. Older systems material stays available as evidence of the path, not nostalgia.
A voice agent latency budget I can actually debug
An NDA-safe field note on where voice-agent latency accumulates, which stage owns each delay, and why interruption recovery belongs in the budget.
Provenance-first agent memory
Why source, trust, sensitivity, allowed use, expiry, and audit gates must remain first-class in an agent memory system.
Static Initialization Order Fiasco
A compact C++ systems note preserved from the old wiki: why global object initialization order leaks across translation units.
The Stack and the Heap
Memory placement as a maker primitive: ownership, lifetime, and why runtime shape matters for C, C++, and Rust.
Indoor Localization with UWB
From the old blog: UWB ranging, local Bluetooth configuration, GPS for outdoor/time sync, and real access-control hardware constraints.
Minimalistic Approach to Cryptography
A cleaned-up cryptography primer: define the terms, keep the threat model visible, avoid ceremony.
Fibonacci coding
A preserved algorithm/wiki reference from the old site: universal integer coding using Fibonacci numbers as a compact representation exercise.
Emacs variants
A tiny old wiki entry promoted into a craft note: editor variants, workflows, and the personal-tooling lineage behind the site.
Local AI Agent Interfaces
Design notes for making autonomous tools feel inspectable: command logs, approval moments, abort buttons, and resumable context.
Architecture, not category
Across the stack on purpose.
My work moves from audio frames and constrained devices to agent orchestration, retrieval policy, telemetry, and production operations. I care about explicit ownership, bounded failure modes, and evidence that survives the demo.
Make invalid states and hidden latency visible.
- Explicit owners over abstraction fog
- Percentiles and traces over dashboard averages
- Fail-open context, fail-closed side effects
- Executable evidence over confident prose
Open a useful thread
Hard systems problem?
I am interested in real-time AI, voice systems, agent infrastructure, memory, observability, and architecture work where the runtime matters as much as the model.