Since I started shipping

Owned the human-calling dialer integration across its whole surface.

• ~3,000 → 30,000 calls/day
• 10× outbound capacity

• Campaigns, bulk lead import, agent provisioning, the calling screen, post-call reconciliation.
• Promoted borrowers to an immediate callback on an intent signal, cutting wasted dials.
• Root-caused a duplicate-dialing incident and a disposition submit that closed the call frame mid-write.

• Spring Boot
• Call-Events
• Webhooks

Automated the lifecycle that keeps a shared number pool dialable.

• ~400-number shared pool
• 4 clients, 0 cross-client collisions

• Spam scoring, provider sync, replacement purchasing, nightly churn.
• Lock-free coordination so three clients on separate databases never burn the same number.
• Fenced off regulated number series after one was destroyed at the provider.

• Java
• Spring Boot
• Schedulers
• Telephony
• Multi-tenancy

Shipped across the speech → LLM → speech path the AI callers run on.

• 600K+ AI calls/day

• Call orchestration and retries, recording, transcription, disposition mapping.
• Dynamic prompt variables — each call carries the borrower's real numbers, spoken as words.
• Per-tenant dashboards, plus hangup cause and answering-machine detection instead of one flat “not connected”.

• Voice AI
• LLMs
• Real-time speech
• Telephony
• Java
• Spring Boot

Moved per-client behaviour out of releases and into a database row.

• ~40 call sites on one seam
• 0 redeploys for config changes

• ~40 existing call sites routed through a single configuration seam.
• Tuning a client ships in minutes instead of waiting on a deploy window.
• Onboarding a client stopped meaning another hardcoded branch.

• Java
• Spring Boot
• PostgreSQL
• Multi-tenancy

Made the jobs behind allocation, prioritisation and retries behave the same way every run.

• Hot-lead prioritisation made deterministic — newest record wins, no mutable column two jobs can race on.
• Optional scheduling inputs made explicit, so “skip” stopped quietly meaning “now”.
• Missing configuration now fails loudly at boot rather than half-working in production.

• Java
• Spring Boot
• Concurrency
• Schedulers
• Sentry

The live view the collections team runs its day on.

• 40+ alerts/10+ Dashboards

• Business Analytics & Alerting
• Live Monitoring
• Real time warning and error alerting

• ClickHouse
• Grafana
• SQL
• Prometheus
• Dashboards-as-code

Turned two years of merge-request feedback into a pre-push checklist.

• 154 MRs analysed
• 123 review comments
• 64 rules distilled

• 123 reviewer comments across 154 MRs distilled into 64 cited rules.
• Every rule quotes the original comment, so it can be argued with rather than obeyed.
• The finding behind it: fixing the flagged line instead of the pattern was buying a second review round.

• Code review
• Static analysis
• Technical writing