Jan 2025 — Present
Sr. Full Stack & AI Engineer @ Swivl
Led development of Swivl's real-time Voice AI Agent and its supporting infrastructure for a self-storage conversational AI platform. Owned the voice pipeline end-to-end — from Twilio telephony ingestion to LLM response — built bi-directional FMS integrations, and stood up the AWS infrastructure and CI/CD serving 4,500+ facilities.
<500ms
p95 end-to-end latency
−40%
live-agent routing QoQ
35%
calls served from cache
- Architected the Voice AI Agent in Python/FastAPI: ingests Twilio Media Streams as base64 μ-law audio, runs Whisper STT and feeds transcripts into an LLM loop with barge-in handling — cutting live-agent routing 40% QoQ at sub-500ms end-to-end latency.
- Designed a Redis semantic caching layer that hashes caller-intent embeddings and serves stored answers for high-frequency questions (gate codes, balances, office hours), bypassing inference on 35% of inbound calls and shaving 600ms off p95.
- Built the intent classification engine with LLM function calling, routing utterances to autonomous actions (payments, access codes, autopay), knowledge-base retrieval, or live-agent escalation with full context summaries.
- Assembled the pgvector RAG system that chunks and indexes facility FAQs, storage policies and unit pricing — grounding every answer in each property's actual data.
- Shipped the Next.js Live Monitor: React Query + SSE stream real-time transcripts, LLM sentiment scores flag negative tonal shifts, and managers toggle which intents the AI handles vs. escalates.
PythonFastAPIWebSocketsNext.jsReactTypeScriptTailwind CSSReact QueryRedisPostgreSQLpgvectorCeleryOpenAI APITwilio Media StreamsWhisperStripe ConnectSSEOAuth2/JWTDockerAWSTerraformGitHub ActionsPytest