One-time challenge replay guard
A local challenge guard that binds principal, action and expiry to one consume decision.
5/5 local tests passed
Engineering / Things I build
What I built. How it works. Where it breaks.
I have built and supported native mobile applications, backend services and security controls used in enterprise environments. I tend to start at the trust boundary, trace the request to the service that owns the decision, and then look at what happens under concurrency, degraded dependencies and a difficult rollout. Publicly shared descriptions, code samples and captured test evidence are separated from private customer systems.
Selected systems I built
These are systems I built and worked on in enterprise environments. I explain the implementation and failure modes, and leave out customer code, internal topology and private operational data.
Hardening iOS Device Trust SignalsI hardened iOS device-trust decisions by separating untrusted client signals from evidence verified by backend services.Session Recovery Without Blind TrustI worked on session-recovery controls that made identity checks, step-up decisions and failure states explicit instead of weakening authentication during recovery.Backend Verification Under LoadI built backend verification safeguards so retries and dependency outages would not quietly bypass risk and authorization checks.
VOICE / AGENT INTEGRATIONS
Five voice-agent projects
I built the workflow APIs, adversarial evaluations, operational models, CRM event processing and native mobile integration interfaces, and documented the trust boundaries and rollout decisions.
SPEECH & AUDIO ML
Audio modeling and inference projects
I implemented a small neural vector-quantized codec, ASR metrics, offline speech synthesis with prosody planning, streaming endpoint detection and controlled ML ablations. The results are measured on clearly described local datasets.
Separate technical test implementations
Full evidence and limitations →16 passing local assertions across three small implementations I built to reproduce failure modes (November 5, 2026). They are not production benchmarks.
Idempotency under concurrent requests
Tenant-scoped operation keys, matching request fingerprints and a shared effect callback.
6/6 local tests passed
Refresh-token rotation under replay
Single-use refresh credentials with token-family generation, revocation and expiry.
5/5 local tests passed