I build AI & software systems for messy real-world workflows— not just polished demos. 

messNo existing AI agent infrastructure. The mandate was to research and prototype from scratch with no prior internal reference.

systemBuilt Singtel Claw, an agent gateway POC that uses Portkey to route prompts to different orchestrator and subagent models based on intent. Spent time evaluating agentic frameworks (LangChain DeepAgents, OpenAI Agents SDK) to find what actually fits the use cases. Proposed a config-first assembly model so teams can spin up new business use cases without touching the orchestration layer each time. Also researched voice agent architecture for B2B deployment, comparing LiveKit, Twilio, and Voltade Envoy, and designed a SIP/PSTN routing strategy using Singtel's IMS core.

• →POC in active validation against real use cases
• →Config-first model proposed to cut per-use-case orchestration work
• →Voice routing strategy ready for B2B deployment evaluation

messProduction clusters had piled up with deprecated LLM deployments and stale configs nobody touched. Monitoring was noisy enough that alerts had lost credibility. No WebSocket support, no load testing, no CI for the proxy.

systemCleaned out deprecated models and configs from the Kubernetes clusters, then tuned vLLM and TRT Engine inference parameters and benchmarked models under real load. Migrated the mock LLM service to FastAPI for WebSocket support, benchmarked WS latency, and designed per-tenant metering across the full chain from Istio ingress through OPA auth to the upstream provider. Built a distributed Locust load testing setup on Kubernetes to validate token limits and session persistence. Wrote client scripts for Azure OpenAI GPT-Realtime and Gemini Live, and set up a GitHub Actions pipeline that runs integration tests on every PR against the WS proxy.

• →+3.6% throughput, -2.3% avg response time at 1k ISL / 200 OSL
• →-6.5% avg response time at 4k ISL / 500 OSL; concurrent user capacity up to 6
• →70% reduction in monitoring noise
• →Rate-limit config moved to SAP HANA DB for centralised management

messDeploys were slow, builds were slower, and production bugs had been accumulating without anyone owning resolution. The CI/CD setup was fragile enough that people avoided touching it.

systemRebuilt the CI/CD pipeline around AWS ECR, EC2, and GitLab Runners. Trimmed the frontend build process. Worked through 10+ production hotfixes across the stack and wrote a handful of complex SQL queries for business and editorial reporting.

• →15% faster deployments
• →20% reduction in build time
• →10+ production bugs fixed, stabilizing a system that had been drifting