Platform and ML Infrastructure Engineer building production AI systems, on 18 years of telecom network performance at Huawei

Multi-agent systems on Google Cloud. Scale-to-zero GPU inference on Kubernetes. Terraform on AWS. Fine-tuned specialist models for edge devices.

Four ADK agents turn a customer complaint into a triaged NOC incident ticket in under 30 seconds

A Google ADK SequentialAgent orchestrates four LlmAgent sub-agents on Gemini (Vertex AI), collapsing the manual NOC workflow of correlating network events, call detail records, and ticketing into one natural-language step. Every LLM call routes through a 4-attempt model-failover ladder shown live in the UI, so a rate-limited model is swapped for the next one on screen instead of failing the request.

• Top 100 of thousands at the Google Cloud Gen AI Academy APAC 2026 (Cohort 1); the BigQuery + AlloyDB pilot now runs zero-idle-cost on bundled SQLite, swappable back through one MCP Toolbox tools.yaml line, live on Cloud Run at netpulse-ui.run.app

Production GPU inference on Kubernetes that costs $0 when idle

Two-layer autoscaling: KEDA watches Redis queue depth for 0-to-N pod scaling while the GKE Cluster Autoscaler provisions and deprovisions GPU VMs on pending pod scheduling. vLLM serves with continuous batching; model weights persist on a PVC and image layers pre-cache via Secondary Boot Disk, cutting cold start about 45% (9m20s to 5m9s).

• TTFT p95 ~140-200ms warm, 12-panel Grafana dashboard (Prometheus + NVIDIA DCGM), Locust load testing, true scale-to-zero at both pod and node level

A 270M model that beats generalist function-callers by baking tool knowledge into its weights

LoRA (r=128) fine-tuning on Gemma 3 270M moves all 14 tool definitions out of the prompt and into the weights: a ~20-token query goes in, a JSON tool call comes out, regardless of tool count. The result is 99.5% accuracy at 32x fewer prompt tokens than schema-in-prompt baselines, in a 291 MB Q8_0 GGUF that runs on phones, laptops, and Raspberry Pi.

• Trained on a single consumer GPU (RTX 4060, 8GB VRAM): 418/420 eval accuracy, 14,033 training examples, 14 tools across 2 MCP servers, 153ms avg latency

The public Terraform and Docker deployment behind a framework-free multi-agent RAG system

AgentLens is a streaming pipeline debugger for multi-agent RAG, coordinating four roles (Retrieval Agent, Grader, Quality Judge, Fallback) across a 4-service FastAPI stack. This repo is the infrastructure-as-code only (app source stays private), so an infra reviewer can verify the full production architecture without app-repo access: Terraform provisions VPC, EC2 t3.small, and Elastic IP in us-east-1, behind a 7-container Docker Compose stack, path-based nginx, and Cloudflare edge SSL.

• One-command deploy and teardown over SSH, Ollama Cloud API (no GPU on EC2), ChromaDB vector store, systemd boot automation, live at agentlens.adityonugroho.com

Four independent detection phases feeding a fault-tolerant orchestrator

Automates construction blueprint takeoffs by composing three independent CV modules (contour-based shape detection, Tesseract OCR with preprocessing, and a YOLOv8n model fine-tuned on synthetic symbols) into a multi-stage orchestrator. Each phase returns a typed failure object instead of aborting, so a first-run user without trained YOLO weights still gets shape and OCR results. Multi-page PDFs go in, structured JSON takeoff reports come out over a FastAPI endpoint.

• 66 passing tests across 4 phases, mAP@50 0.992 on 5 synthetic construction symbol classes, YOLOv8n fine-tuned 20 epochs, FastAPI POST /analyze plus CLI

Six end-to-end telecom ML projects, each on synthetic data with embedded network physics

Six self-contained projects spanning classification, regression, time-series, and reinforcement learning. Most ML demos treat telecom as generic tabular data; here each project hand-crafts synthetic data with embedded telecom physics (temporal correlation, spatial clustering, equipment failure signatures), then runs the full paradigm end-to-end from data generation through feature engineering, training, evaluation, and business-insight translation. The repo is an index; source lives in six child repos.

• Churn AUROC 0.86, RCA Acc@1 0.91, Anomaly F1 0.70, QoE RMSE 0.45, Capacity MAPE 14.5%, Network Optimization +61% vs random

• Location: Jakarta, Indonesia (UTC+7), open to remote
• LinkedIn: linkedin.com/in/adityonugrohoid
• Email: adityo.nugroho.id@gmail.com