Shan Wijenayaka.

Lead AI Engineer

Certis · AI Platform & Security Operations

Leads the 5-engineer team behind the enterprise agentic AI platform; 8 enterprise deployments including a major transport hub; 35% of tier-1 support tickets deflected.

An operational orchestrator for security operations, acting on inputs from voice, vision, and humans in the loop. Manage the 5-engineer team: ran 60+ interviews to build it, and designed the 6-week onboarding curriculum that converts strong backend and data science hires into production-AI engineers. Own the platform's 99.9% SLO; built its call center agent end to end; now in beta handling inquiry and case-management calls ahead of production go-live. RFC/ADR process adopted by two sibling teams; represent the platform on two customer AI steering committees.

• ·Orchestration core and MCP connectors, built with the team. LangGraph agents with shared state, handoff contracts, and planning loops on Temporal, chosen over event choreography for replayable, compensable writes. First four MCP connectors (ITSM, video management, access control, workforce) delivered as a team, scaling to 14 access-scoped connectors behind one audited interface, with A2A interop. LiteLLM tiered routing: self-hosted vLLM (Qwen fast default, SEA-LION for reasoning-heavy turns) with Claude on Bedrock and OpenAI as managed fallback, for a cost ceiling, data residency, and graceful degradation; one CI/CD pipeline, MLflow lineage.
• ·Voice AI at concurrency. Duplex SIP/WebRTC through LiveKit with barge-in and Silero VAD, streaming ASR/TTS, and a per-stage latency budget across STT, retrieval, model, and TTS. p95 turn latency 480ms at 60 concurrent sessions; 45k calls/month in English, Mandarin, Malay, and Tamil; 82% containment without human handoff. Inference spend down 28% via continuous batching, KV/prefix caching, AWQ quantization, and Redis-backed semantic caching on the RAG hot path.
• ·Vision with VLM grounding. YOLO-family detectors (PyTorch) across 240 in-house camera streams on Kafka at 6 sites, TensorRT INT8 at the edge; Qwen3-VL and Gemma 4 26B-A4B-it grounding fires only on detector hits, amortizing VLM cost. Active learning and hard-negative mining cut false positives 63%; operator-confirmed precision 0.91. Vision events enter the agent loop as tools across lost-and-found, hazard recognition, vehicle and parking, and occupancy workflows.
• ·Operator console, end to end. React/Next.js with streaming output, write-confirmation UX, WebSocket state sync, WebRTC push-to-talk, Playwright e2e. 300+ of ~380 eligible weekly operators; handle time down 22%.
• ·Safety architecture and incident command. Sandboxed tool execution with egress controls, blast-radius limits, human confirmation, automated rollback on KPI degradation; quarterly red teaming; controls mapped to the IMDA framework ahead of customer audits. Zero unapproved writes; unsafe tool calls down 89%. Incident commander for the platform's only Sev-2; the postmortem set org-standard backoff budgets and a weekly ops review.
• ·Evaluation and tuning flywheel. Golden sets with ops SMEs, LLM-as-Judge on ~2k weekly episodes in Langfuse, benchmark in CI, per-language WER and latency budgets for voice, canary and interleaved online evals, DPO on operator accept/reject pairs, and a LoRA fine-tune on three months of operator-call recordings that surfaces the right SOP as a call unfolds. Task success 78% to 91%; failure taxonomy maintained on the residual 9%.