BLR11:02:10
···05:32:10
00:00
X0.0000
Y0.0000
← Back to projects
Client Delivery

Conversational Analytics - HSBC

HSBCHSBC
GIL'd threading caused memory ballooning - concurrent sessions saturated at 20 per VM before packet loss rose above 10%; available hardware capacity was highly under-utilised; post-call documentation required 10–15 minutes of manual effort per interaction with no automated path; fragmented cross-service logs with no correlation layer meant incidents required 1–2 hours of manual reconstruction to identify root cause.
⚠️The platform ran on GIL'd threading with memory ballooning on a 32-core VM - capped at 20 concurrent calls, 31 cores idle. Post-call documentation: 10–15 min per interaction. Inference cost: ~$118K/month. Incident recovery: 1–2 hours - fragmented logs, no unified observability layer.
⚙️Led a 4-engineer team. Owned Packer automation across all project modules - standardizing GCE image builds for the full SIP stack (SBC → STT → LLM inference). Migrated from 32-core n2-standard-32 to 8-core c4-standard-8; deployed 8 CPU-pinned parallel instances via <code>taskset</code> - one per core, escaping GIL entirely. Rewrote concurrency with asyncio + uvloop, eliminating memory ballooning. Built SIPp load test suite (2,000 concurrent users). Architected cross-stack log-correlation over GCP Logging APIs - 250K+ log lines in under 5 seconds.
🛡️<2s E2E transcription latency, <5% packet loss at 1,600+ concurrent sessions. libsrtp + DTLS/SRTP for in-transit security. Grafana-Prometheus with MACD triggers. Migrated n2-standard-32 → c4-standard-8, improving transcript length 30–40% under load.
🚀7× per-VM capacity (20 → 140–160 calls). 1,600+ sessions sustained. Documentation: 10–15 min → 2–3 min. Compute: $118K → $8K/month (~$1.3M annualized savings). MTTR: 1–2 hours → ~10 minutes.
1,600+ concurrent sessions • 7× VM capacity • ~$1.3M annualized savings • MTTR ~1–2 hrs → ~10 min
SIP/Voice Orchestration (PJSIP • RFC 3261)CPU-Pinned Processes • asyncio + uvloopGCP Infrastructure (Packer • GCE • HPA)Cross-Stack ObservabilityLLM Inference PipelineCost Engineering