Projects¶
A portfolio of production-grade AI systems, from physics-informed energy platforms to agentic RAG chatbots, MLOps pipelines, and developer tools.
Ongoing: Enterprise AI Automation¶
Agentic ERP Automation Platform LIVE IN PRODUCTION¶
An agentic automation platform layered on a live commercial cloud ERP, running the whole supply chain of a multi-entity food manufacturing group: buy, receive, make, sell to the group's own outlets, and pay. Purchase orders, partial goods receipts, AI vision extraction of vendor invoices matched to their receipts, production and work orders from the real bill of materials, an inter-company sales flow from outlet spreadsheet to settlement, stock opname and transfers from a Flutter Android app, and a three-gate payment flow with enforced separation of duties that executes real bank transfers and writes payments back into the ERP.
Built with no ERP sandbox and a hard daily API quota, so every safety property had to be engineered: a GET-only mirror for reads, a seven-layer write-guard with idempotency and read-back verification, timeouts treated as unknown outcomes rather than failures, a shared quota guard across 20 services, and per-capability feature flags canaried one document at a time.
Technologies: Python, FastAPI, Next.js, TypeScript, Flutter, PostgreSQL, Claude (headless CLI, vision, MCP), Docker Compose, Prometheus/Grafana, GitHub Actions
Highlights: 5 operational pillars, 12 guarded ERP write capabilities, ~310 API endpoints, 61 read-only AI tools, 620 automated tests, 75 tables, 1,485 commits, live and moving real money
AI Drawing Platform for Architecture and Interior Design IN USE¶
An internal platform that turns a designer's chat into exact working drawings: floor, finish, ceiling and lighting, electrical and plumbing plans, sections and elevations, a door and window schedule and a quantity take-off, plus a 3D model and a browser walkthrough of the same space. The model never draws: Claude edits a validated JSON spec and deterministic engines render every sheet to exact scale, so every output agrees with every other.
Proved on a real outlet's 15-page drawing set, transcribed and checked by its own clash checker. Sections are cut from the stacked 3D model and re-cut when a floor changes; electrical, plumbing, finish and ceiling sheets are live views of their floor; door and window codes stay stable across revisions.
Technologies: Python, ezdxf, trimesh, FastAPI, three.js, Claude (headless CLI, vision), Blender Cycles, Docker Compose
Highlights: one spec drives every sheet, the quantities and the 3D walkthrough; ~9 s chat edits; PDF import in ~16 s; clash checking fed back to the model
Ongoing: AI Game Studio¶
AI Game Studio Automation for Mimio Entertainment ONGOING¶
An AI automation system that runs a small Roblox game studio from research to publishing. Twelve role agents on headless Claude (research, compliance, design, art direction, Roblox and Unity builders, QA, store, marketing, content, analytics) work through an 18-step launch pipeline; the developer approves or sends work back with notes, and builds the game in Roblox Studio through the Studio MCP. Roblox Open Cloud handles game passes, data checks and live-server tests; a web dashboard and an Obsidian vault show every step.
The first game, Wisteria Hollow, ships with saved progress and a cross-server session lock, Bahasa Indonesia and English, touch controls and four fair game passes, and was renamed and reworked after an AI review flagged names too close to an existing anime.
Technologies: Claude (headless CLI, Opus), Roblox Studio MCP, Luau, Roblox Open Cloud, Python, FastAPI, Docker Compose, YouTube Data API, Obsidian, ffmpeg
Highlights: 12 agents, 18-step pipeline with approval gates, game built and published on Roblox, store art and a YouTube Short produced by the pipeline, engine-agnostic for Unity
Mobile & Kiosk Apps¶
Face-Attendance Android App + Outlet Kiosk FLUTTER LIVE¶
One Flutter codebase shipping as two products: an employee phone app where clocking in is a face verification inside a geofence, and the same binary running as a shared outlet kiosk tablet with no login and no session. The company's attendance clock since October 2026, used every day on outlet tablets.
Built for store floors rather than for a demo: a server-issued liveness challenge with server-tuned coaching limits and a three-frame burst verified server-side, no face embedding ever computed on the device, an offline clock queue replayed with its original timestamp, and NIK + PIN sign-in with no per-login message cost. The kiosk is locked down by a native Kotlin layer (launcher alias, lock task, device owner) that replaced a paid kiosk app, installs its own updates silently, and reopens itself after an update or a power cut, including the Android 15 overlay workaround.
Technologies: Flutter, Dart, Kotlin, Riverpod, go_router, dio, Google ML Kit face detection, camera, geolocator, flutter_secure_storage, video_player, wakelock_plus
Highlights: in daily production use, ~7,600 lines of Dart, over 35 over-the-air releases in the first month, tablet clock cut from about a minute to seconds, self-recovering kiosk tablets, roster-scoped matching with a NIK fallback
Ongoing: Workforce Platforms¶
HR & Face-Attendance Platform LIVE¶
A company-owned HR platform for a multi-site food and beverage business, built on three surfaces over one backend: an Android app where employees clock in by face inside a geofence, the same app running as a locked-down kiosk tablet in the outlets, and an HR web dashboard for attendance, contracts, incidents, salary, recruitment with psychological tests, reimbursement and cash advances, and payroll. The company's attendance system since October 2026.
Face verification runs server-side on CPU (InsightFace ArcFace) with a review band instead of a guess between the accept and reject thresholds, a kiosk that matches only against its own sites' roster for the day, and adaptive templates learned only from unambiguous matches. An attendance rules engine handles night shifts across midnight, breaks, overtime and staff who move between branches. Contract recommendations come from a rule engine with the evidence shown; the model writes the narrative and never produces a number. Biometric handling is designed against Indonesia's UU PDP No. 27/2022.
Technologies: FastAPI, PostgreSQL + PostGIS, Next.js, Flutter, InsightFace + onnxruntime, Claude (headless CLI), MCP, Docker Compose, Prometheus, GitHub Actions
Highlights: dozens of employees clocking by face daily, first-try match 72% to 90% after adaptive templates, ~320 API endpoints, 71 tables, 22 read-only AI tools, 225 backend tests, several hundred employees and dozens of sites on the platform, payroll pending a parallel run
Ongoing: AI-Native Applications¶
Two applications built on the same architectural idea: one shared live state edited by both a human and an AI agent, an MCP tool layer instead of code generation, and a headless Claude Code agent running on a subscription.
TransForce - AI-Native Particle Transport DEPLOYED¶
Chat-driven Monte Carlo radiation transport on OpenMC. Describe a reactor physics problem in plain language and an agent builds the CSG geometry, materials, source and tallies, runs the transport calculation, and reports k-eff with uncertainty. Ingests a paper from an arXiv id, DOI, URL or PDF, extracts a structured model spec, reproduces it, and reports the bias in pcm against the published values.
The agent never writes Python: it manipulates a typed, provider-agnostic ModelSpec, and a deterministic driver translates that into live OpenMC objects. Includes depletion, weight-window variance reduction, MGXS generation, and PDF reporting.
Technologies: Python, FastAPI, FastMCP, OpenMC, Pydantic, React, Vite, matplotlib, reportlab, Docker
Highlights: 27 simulation MCP tools, 4 benchmark presets including Godiva, 3 compute backends (local/VPS/GCP), C/E comparison in pcm, runs on a Claude subscription with no API key
CutForce - AI-Native Video Editor DEPLOYED¶
A browser video editor where a human and an AI agent edit the same timeline simultaneously, over 43 MCP tools and a live WebSocket-synced store, with an embedded terminal wired to the editor's own MCP endpoint. Undo works across both editors because every mutation funnels through one commit path.
The export engine compiles the timeline into a single ffmpeg filter_complex: keyframes become piecewise-linear ffmpeg expressions, speed changes chain atempo stages, and colour grades resolve from the same numbers in both the CSS preview and the ffmpeg render.
Technologies: TypeScript, Express, MCP SDK, React, Vite, zustand, xterm.js, ffmpeg, node-pty, Veo/Imagen/FLUX, Docker
Highlights: 43 MCP tools, 7 authored agent skills, 12 grade presets, cross-editor undo, generative clips with cost estimation, runs anywhere Node and ffmpeg run
Flagship — ForceX AI¶
ForceX AI — AI-for-Energy Platform¶
Indonesia's first AI-for-energy platform — 12 physics-informed AI products for nuclear, geothermal, oil & gas, and renewables. Multi-agent orchestration with LangGraph, CNN/PINN/GNN/LSTM models, and production deployment.
Technologies: PyTorch, LangGraph, PINNs, GNNs, CNN, LSTM, RL (PPO), FastAPI, React, Docker
Highlights: 12 products, 484 tests, GeoForce deployed (R²=0.997), published on HuggingFace
View Project → | Visit forcex-ai.com →
GeoForce — CNN Geothermal Reservoir Surrogate¶
Standalone production version of the GeoForce CNN surrogate model. Replaces hours of TOUGH2 geothermal simulation with sub-second inference at R²=0.997 accuracy.
Technologies: PyTorch, CNN, TOUGH2, FastAPI, HuggingFace Hub
Highlights: 57K params, R²=0.997, model + dataset published on HuggingFace
LLM Fine-Tuning & Alignment¶
EnergyLM-7B — LLM Fine-Tuning & Alignment Pipeline IN PROGRESS¶
End-to-end LLM training pipeline for a domain-adapted energy model. Fine-tunes Qwen2.5-7B using QLoRA SFT on 20K synthetic instructions, compares DPO vs ORPO alignment, trains a reward model, and evaluates across 10 benchmarks. Multi-teacher data generation, dedup/filter pipeline, AWQ/GGUF quantization, vLLM serving. $0 budget — 100% free compute.
Technologies: PyTorch, Transformers, TRL, PEFT, QLoRA, DPO, ORPO, lm-eval-harness, vLLM, AutoAWQ, llama.cpp, Kaggle T4
Highlights: 20K synthetic dataset, DPO vs ORPO comparison, 10-benchmark evaluation, CoT distillation, reward modeling, AWQ + GGUF quantization, $0 compute cost
View Project → | ForceX-AI on HuggingFace →
GenAI & LLM Systems¶
Enterprise Agentic RAG Chatbot¶
Agentic RAG system with LangGraph multi-agent orchestration, self-reflective retrieval, hybrid search, and streaming responses. Deployed in production for enterprise documentation Q&A.
Technologies: LangGraph, LangChain, FastAPI, React, pgvector, Langfuse, Prometheus
Highlights: Self-reflection loop, 14 specialized agents, multilingual (ID + EN), production-deployed
Agentic HR Intelligence Platform¶
Full-stack AI-powered HR platform that transforms raw employee and attendance data into workforce intelligence. Features an agentic AI chat with 5 specialized tools (Gemini 2.5 Flash), interactive analytics dashboards, turnover risk scoring, anomaly detection, and AI-driven contract renewal recommendations.
Technologies: Next.js 16, React 19, FastAPI, PostgreSQL, Google Gemini 2.5 Flash, Recharts, Tailwind CSS 4
Highlights: 5 agent tools with 10-iteration tool calling, 5-factor risk scoring (0-100), 5-type anomaly detection, bilingual (ID + EN), PDF report generation
AI-Powered Help & QnA System¶
Production-ready RAG system with hybrid retrieval (dense + BM25 + RRF fusion), multi-provider LLM fallback chain (5 providers), cross-encoder reranking, and voice assistant.
Technologies: FastAPI, pgvector, BGE embeddings, Redis, React, Docker, Kubernetes
Highlights: 40K+ docs indexed, 75%+ cache hit rate, ≤5% hallucination rate, 100% free-tier LLMs
Porto Agent: LinkedIn Drafts From Real Work¶
Turns real engineering work into first-person LinkedIn drafts, then refuses to post any of them until I say so. Stories come from anonymised portfolio pages and short notes, not commits; Claude (headless CLI, tools disabled) writes one concrete moment per post in text, diagram or carousel format; a content guard fixes or blocks every draft; and publishing is a separate, manual, server-gated step.
Technologies: Claude (headless CLI), FastAPI, Next.js, PostgreSQL, LinkedIn Posts API, ReportLab, nginx auth_request, Firebase Auth
Highlights: honest evaluation of AI-sounding drafts and the fixes, a guard that blocks private terms and invented details, publishing that cannot happen by accident
Computer Vision & Video Analytics¶
Retail Video Analytics - Customer Behaviour from CCTV COMPLETED¶
An end-to-end pipeline that turns two retail CCTV cameras into behavioural metrics: store interest and walk-in conversion, per-shelf engagement events, and genuine staff-to-customer interaction sessions. YOLO11m detection with ByteTrack identity persistence, entirely on CPU, reproducible from one Docker command.
The engineering is in the layer above the detector. Perception is cached so that behavioural thresholds can be tuned interactively instead of at 80 CPU-minutes per iteration. Speeds and distances are normalised by body height so a single threshold holds across the depth of the frame. Shelf assignment uses the torso rather than the feet, after an audit showed browsing customers' feet stay in the aisle while their upper body leans into the fixture. An apron-colour staff classifier was built, measured, and rejected in favour of a spatial rule that actually separates the classes.
Technologies: Python, YOLO11m (Ultralytics), ByteTrack, PyTorch CPU, OpenCV, pandas, YAML, Docker
Highlights: 2 cameras, 3 behavioural metrics, 10 fps effective inference on CPU, zone geometry validated against plotted real tracks, interpolation guard that stopped a phantom interaction, documented misses
AI Applications¶
Deployable AI product builds, each a live, working demo grounded on Gemini 2.5 Flash (Vertex AI), built with Next.js and deployed on a VPS with automatic HTTPS.
AI Sales & Lead-Gen Chatbot (WhatsApp / Instagram)¶
A 24/7 sales assistant that answers pricing/schedule questions, handles objections, captures every lead, classifies status, and saves to a CRM with hot-lead alerts. Web chat plus WhatsApp/Instagram channels.
Technologies: Gemini 2.5 Flash, Next.js, JSON-mode extraction, file CRM, Telegram alerts
Highlights: Two-call architecture (reply + structured lead extraction), status classification, live admin dashboard
Automated LinkedIn Carousel Generator (RAG)¶
Upload documents and a RAG pipeline writes a fresh 8-10 slide LinkedIn carousel grounded in the content, with a topic-dedup log, exportable as a LinkedIn-ready PDF.
Technologies: Gemini 2.5 Flash, Vertex text-embedding-004, vector retrieval, jsPDF + html-to-image, Next.js
Highlights: Grounded slide generation, topic dedup, native carousel rendering (no Canva API), isolated PDF parsing
Self-Correcting Document Pipeline (LangGraph)¶
A supervisor-worker LangGraph agent that extracts invoice data, validates the arithmetic deterministically, self-corrects on mismatches (capped retries), and writes a report — with a live graph, streamed steps, and an eval panel.
Technologies: LangGraph (@langchain/langgraph), Gemini 2.5 Flash, React Flow, SSE, Next.js
Highlights: Real StateGraph with conditional self-correction loop, deterministic validation, per-run token/latency/loop evaluation
AI Assistant Embedded on a Portfolio Site¶
A floating AI chatbot embedded on a personal site that answers visitors' free-form questions about skills, projects and availability, grounded in the site content so it never invents facts.
Technologies: Gemini 2.5 Flash, Next.js, knowledge-base grounding, embeddable widget
Highlights: Anti-hallucination grounding, hire-intent nudging, drops into any existing site
WhatsApp Lead Bot with Photo-on-Demand (Real Estate)¶
An always-on WhatsApp assistant for studio rentals that answers questions, sends a property photo gallery on request, captures leads, and books viewings, with editable galleries.
Technologies: Gemini 2.5 Flash, Next.js, token-driven photo delivery, file CRM
Highlights: Text-plus-photos via model-emitted tokens, grounded answers, lead capture and status
WhatsApp AI Chatbot with Product Compatibility (Automotive)¶
A WhatsApp chatbot for an automotive electronics business: KB-grounded answers, product compatibility by car model with upgrade suggestions, rule-based human handoff, and lead capture.
Technologies: Gemini 2.5 Flash, Next.js, compatibility rules + KB grounding, file CRM
Highlights: Model-aware product recommendation, deterministic premium-brand handoff, editable knowledge base and rules
InsightFlow -- AI Analytics Workspace¶
Upload a CSV and the workspace auto-builds a dashboard: KPIs, anomaly detection across every metric, a trend chart, AI-written insights, and a natural-language layer to ask questions about the data. Deterministic stats in code, LLM for narration only.
Technologies: Gemini 2.5 Flash, Next.js, Recharts, PapaParse, TypeScript analytics engine
Highlights: IQR anomaly detection, grounded AI insights and Q&A (no hallucinated numbers), BI-style workspace UI
MarketMind -- AI Crypto & Stock Market Intelligence¶
A market-intelligence SaaS: live crypto prices, technical indicators (SMA, RSI, volatility), a Bullish/Bearish/Neutral signal, AI-written analysis grounded on the indicators, and a chat to ask about any asset. Polished dark trading terminal.
Technologies: Gemini 2.5 Flash, Next.js, Recharts, CoinGecko, TypeScript indicators engine
Highlights: Live data, technical analysis in code, grounded AI analysis + Q&A (no hallucinated numbers), signal generation
DataPilot -- Private AI Data Analyst¶
A self-hostable "code interpreter": describe an analysis in plain English, the AI writes the Python, runs it in a sandbox, and returns a report (charts + results + the code), refined by prompting. Privacy-first -- the model only sees the schema, and it runs on a local model (Ollama) so data never leaves the machine.
Technologies: Next.js, Gemini / local Ollama (Qwen2.5-Coder), Python sandbox (pandas, matplotlib, scikit-learn), Docker
Highlights: Live streaming code-gen, self-correction loop, switchable cloud/local model, sandboxed execution, fully on-premise/air-gappable
HealthPredict -- Clinical Risk Prediction¶
A self-hostable web app that turns routine clinical values into an explained ML risk estimate for common conditions (diabetes, heart, kidney, liver, thyroid). Each condition is scored by a Logistic Regression + Random Forest ensemble, with the contributing factors ranked on every result.
Technologies: Python, Flask, scikit-learn (LogReg + Random Forest), SQLAlchemy (SQLite/MySQL), gunicorn + nginx
Highlights: Two-model ensemble per condition, explainable factor breakdown, history dashboard, admin console, light/dark, production-deployed
Web & Business Systems¶
LeadFlow -- Sales & Lead Management System¶
A full-stack CRM: multi-source lead capture, a visual deal pipeline (New -> Contacted -> Quoted -> Won/Lost) with per-stage value totals, overdue follow-up reminders, and conversion analytics. REST API with a swappable data store.
Technologies: Next.js, TypeScript, Tailwind, Recharts, REST API, file/Airtable/Supabase/MySQL store
Highlights: Kanban pipeline, multi-source capture form, follow-up reminders, funnel + value-by-stage analytics
MLOps & Infrastructure¶
Broadcast Analytics — MLOps Platform¶
Enterprise MLOps platform with 24 trained models (6 algorithms × 4 targets), LLM-powered analytics chat, genetic algorithm schedule optimization, and full monitoring stack.
Technologies: FastAPI, scikit-learn, XGBoost, CatBoost, LangChain, Redis, Prometheus, Grafana, Docker, Kubernetes, GitHub Actions
Highlights: 24 models, genetic algorithm optimizer, Prometheus + Grafana + Alertmanager + Loki monitoring, full CI/CD
Media Platform — Ad Campaign Management¶
Comprehensive enterprise media platform with OTT streaming, broadcast management, media asset management, and multimodal AI (NLP + Vision + Audio). Microservices architecture deployed on Google Cloud Platform.
Technologies: NestJS, Next.js 14, Kafka, PostgreSQL + pgvector, MinIO, CLIP, Whisper, spaCy, Docker, Kubernetes, GCP (Cloud Run, Cloud SQL, Artifact Registry, Secret Manager)
Highlights: 5 Cloud Run microservices in production, managed PostgreSQL with private networking, MediaMTX streaming on Compute Engine, 350+ pages documentation
GPU Computing¶
AI Service Migration: CPU to GPU (RTX 5060)¶
Migrated a multimodal AI inference pipeline (Whisper, CLIP, YOLO) from CPU-only VMs to a dedicated NVIDIA RTX 5060 GPU server. Solved Blackwell sm_120 compatibility issues, managed VRAM constraints on 8 GB, and achieved zero-downtime rolling cutover.
Technologies: NVIDIA CUDA, PyTorch 2.11.0+cu128, faster-whisper (CTranslate2), CLIP, YOLOv8, Docker + nvidia-container-toolkit, Bull/Redis
Highlights: 8.7x pipeline speedup, Blackwell sm_120 CUDA fix, VRAM budgeting (67% of 8 GB), zero-downtime rolling migration with instant rollback
Developer Tools¶
Automation Testing Platform (ATP)¶
Playwright-like testing framework specialized for AI platforms — LLM APIs, AI chat UIs, and intelligent applications. Published on npm with full TypeScript support.
Technologies: TypeScript, Node.js, Playwright, Fastify, Next.js, pnpm + Turborepo
Highlights: Published on npm (@robi-atp/*), semantic similarity testing, hallucination detection, AI-aware browser selectors
Research¶
Recommendation Systems Research¶
Comprehensive comparison of YouTube Two-Tower, Netflix Foundation, and Hybrid recommendation architectures with 108 hyperparameter configurations tested.
Technologies: PyTorch, FAISS, Transformers, Gradio
Highlights: 0.744 best combined score, 3 architectures, interactive Gradio apps
Project Categories at a Glance¶
| Category | Projects | Key Skills |
|---|---|---|
| Mobile & Kiosk Apps | Face-Attendance App + Outlet Kiosk | Flutter, Dart, Kotlin, ML Kit face detection, camera, geolocation, offline sync queue, device-owner kiosk lock, out-of-store releases |
| Enterprise Automation | Agentic ERP Platform, HR & Face-Attendance Platform | FastAPI, Next.js, Flutter, guarded ERP write-back, MCP tool servers, RBAC and audit, Docker Compose |
| LLM Fine-Tuning | EnergyLM-7B | QLoRA SFT, DPO, ORPO, CoT distillation, reward modeling, quantization |
| Physics-Informed ML | ForceX AI, GeoForce | PINNs, CNN surrogates, simulation, HuggingFace |
| GenAI & LLM | Agentic RAG, HR Intelligence, Help QnA, Porto Agent | LangGraph, RAG, multi-agent, agentic tool calling, LLM fallback chains |
| MLOps | Broadcast Analytics, Media Platform | ML pipelines, model registry, CI/CD, monitoring |
| GPU Computing | AI Service Migration (CPU to GPU) | NVIDIA CUDA, PyTorch cu128, Docker GPU, VRAM budgeting, Blackwell sm_120 |
| Developer Tools | ATP | npm publishing, TypeScript frameworks, AI testing |
| Cloud & Infrastructure | Media Platform | GCP Cloud Run, Cloud SQL, Kubernetes, Terraform |
| Research | Recommendation Systems | PyTorch, FAISS, Transformers |