╔═════════════════════════════════════════════════════════════════════════════════════════════════════╗
║  SYSTEM        YOEL MANOPPO // Y0EL                                                                 ║
║  ROLE          AI FORWARD DEPLOYED ENGINEER                                                         ║
║  OPS           AI INFRASTRUCTURE                                                                    ║
║  TEAM          4 CORE ENGINEERS                                                                     ║
║  LOCATION      JAKARTA, INDONESIA                                                                   ║
║  STATUS        ████████████████████  ONLINE                                                         ║
╚═════════════════════════════════════════════════════════════════════════════════════════════════════╝

AI Forward Deployed Engineer with a background in E-Commerce AI Systems and deep self study in algorithmic behavior, signal analysis, and agentic architecture.

Part of a 4 person core engineering team building national scale multi agent AI for BNPT Badan Nasional Penanggulangan Terorisme. This is not a side project. It is production infrastructure running at national scale for intelligence operations.

I do not just wire APIs together. I build systems that reason autonomously across multi tier search pipelines, resolve entity identity through constraint intersection, operate 24 hours without supervision, and scale from 5 workers to 50 on demand.

My thinking crosses domains from OSINT geolocation to local LLM voice pipelines to agentic crawlers to blockchain AI validation. The consistent pattern is systems thinking applied to black box problems where the answer is never on the surface.

[ACTIVE]  BNPT Agentic AI Infrastructure
          National scale multi agent system
          4 core engineers
          Production agentic architecture for intelligence grade operations

[RUNNING] AutoCrawl Overnight Run 2026-05-07
          929 vendors indexed / 12x estimated throughput
          104 fully enriched / 497 translated to Bahasa Indonesia
          Peak throughput 150 vendors per hour / 50 parallel workers

[ONLINE]  ATLAS Intelligence
          Local voice AI pipeline / UIX system
          Whisper ASR + Ollama + edge-TTS / GPU resident

TYPE     Agentic 24hr crawler / Defense and security intelligence
STACK    LangGraph + LangChain + Playwright + Crawl4AI + Ollama (IBM Granite 4.1)
         FastAPI + PostgreSQL + Redis Streams + ChromaDB + Langfuse + Grafana
         Vue 3 + TypeScript + Apache ECharts + MapLibre GL + Vue Flow
INFRA    12 Docker services / GPU overlay optional / Zero cloud dependency default

LangGraph state machine fans out to 50 parallel workers discovering security and defense expo vendors across the open web. Multi tier search pipeline: Wikipedia REST direct, DuckDuckGo, Google News RSS, and OpenSERP running Google, Bing, Yandex, and Baidu simultaneously via headless Chromium.

PDF brosur discovery and OCR via Ollama vision model gemma4:e4b. Vendor identity resolution through schema.org, anchor link analysis, and LLM tiebreak. Chroma vector store deduplicates vendors across runs using cosine similarity. NLLB 200 translates all enriched data to Bahasa Indonesia locally.

Vue 3 tactical dashboard includes a 2.5D world map with cylinder bars per country, fly arcs between intelligence hubs, interactive drilldown panels, and a live LangGraph canvas showing the state machine in real time.

Full observability stack: Prometheus metrics, Grafana dashboards, Langfuse LLM tracing, structlog structured logging, and per domain Redis rate limiting.

RESULT   929 vendors indexed in one overnight run
         12x original throughput estimate
         150 vendors per hour at peak
         104 fully enriched with contacts and addresses
         497 descriptions translated to Bahasa Indonesia

TYPE     Full local voice AI pipeline for UIX intelligence system
STACK    Whisper ASR + Ollama + Groq + OpenRouter + edge-TTS
         FastAPI + WebSocket + CUDA 12.x
INFRA    Runs entirely on device / No cloud required / GPU resident

End to end local voice AI system. Whisper handles speech to text with GPU acceleration via CUDA 12.x. Supports Blackwell RTX 50xx via nightly PyTorch builds. LLM inference routes to Ollama locally, Groq for cloud speed, or OpenRouter for model flexibility. edge-TTS converts responses to natural speech.

WebSocket live mode for real time conversation. Persistent memory layer across sessions. Model benchmark suite compares latency and quality across providers. Built for the UIX intelligence system. Designed to run with zero network dependency.

FEATURES WebSocket live mode / Persistent memory / GPU accelerated ASR
         3 LLM backends / Model benchmarking / Fully offline capable

TYPE     Visual signal to coordinate geolocation system
STACK    Python + OSM Nominatim + Overpass API + GPT-4o + LangChain + Mapbox
INFRA    100 percent free APIs / No paid services required

Geolocates images from visual cues only. No GPS metadata required. License plate prefix mapping narrows search to regional bounding box. Road classification, lane count, and median presence filter candidates from Overpass API. Fuzzy landmark normalization handles misspelled or partial POI mentions.

Constraint intersection logic progressively eliminates candidates that fail multiple criteria simultaneously. Weighted confidence scoring estimates both the most likely location and the radius of error. Outputs text report, GeoJSON, and interactive Folium map for each analysis.

METHODOLOGY  Signal extraction > area refinement > Overpass query
             Constraint filtering > confidence scoring > output
OUTPUTS      Text report / GeoJSON / Interactive Folium map

TYPE     Behavioral signal analysis for fraud detection
STACK    Python / Heuristic scoring / Pattern variance analysis

Detects inauthentic activity through behavioral pattern analysis. Scoring system examines transaction patterns and interaction data for variance signatures that distinguish organic human behavior from synthetic or bot generated activity. Built from first principles understanding of how algorithmic platforms detect and respond to inauthentic signals.

TYPE     VeChain blockchain app with AI image validation
STACK    React + VeChain SDK + Solidity + Hardhat + GPT-4o
INFRA    VeWorld exclusive / Thor blockchain network

Rewards users with tokens for uploading photos of healthy juice drinks. GPT-4o validates image authenticity before triggering smart contract reward. Solidity contracts deployed on VeChain Thor. Full monorepo with React frontend, Express TypeScript backend, and Hardhat contract tooling.

Open to:  AI systems consulting  /  Agentic architecture  /  Forward deployed engineering roles
Region:   Indonesia  /  Remote Asia Pacific  /  International