Prooven Platform

Overview

Prooven is the last-mile operations platform used by ABC Express to run government and enterprise distribution programs — IFP, Pusbuk, Kemensos — delivering tens of thousands of shipments to schools and institutions across Indonesia. It closes the loop between the upstream Antero logistics system and physical delivery: couriers scan TTK waybills in an offline-first PWA, capture proof-of-delivery photos and serial-number verification, and backoffice staff process BAST handover documents into billing — all over a Supabase/PostgreSQL data layer with row-level security, a Go edge gateway, a Rust knowledge-graph + Monte Carlo engine, and a DeepAgents.js AI agent for operational decision support.

Problem

Last-mile delivery for large distribution programs was hard to verify end-to-end: proof of delivery lived on paper, BAST documents drove billing manually, field couriers worked in low-connectivity regions, and clients had no trustworthy read-only view into delivery progress.

Solution

A centralized verification platform spanning field capture to finance: an offline-first courier PWA that survives dead zones, a backoffice dashboard for SLA monitoring, BAST processing, and billing (Pra-PI → PI → SI → Lunas), client-facing view-only access scoped by RLS, and an external JSON:API for program stakeholders — plus a Rust simulation/graph layer and AI agent for forecasting and analysis.

Responsibilities

• Built and maintained modules across the delivery lifecycle — scan, POD capture, BAST, billing
• Designed the offline-first PWA sync layer — IndexedDB outbox, compressed photo pipeline, background retry
• Implemented RBAC + client view-only access with Postgres RLS and security-definer RPCs
• Built the AI operations agent — DeepAgents.js/LangGraph with tools over the knowledge graph, Monte Carlo engine, and shipment data
• Integrated with the upstream Antero platform — TTK lookup and last-mile status/POD push
• Kept quality gates green — Vitest, Go/Rust/Bun test suites, migration lint, Trivy/gosec CI scans

Key Features

• Offline-first field PWA — QR/barcode TTK scanning (html5-qrcode + Zebra SDK), serial number box/body verification, 8-photo POD evidence, GPS geostamps, IndexedDB outbox → auto-sync, packaged via Capacitor for Android
• BAST pipeline — scan intake, OCR assist, revision workflow, promotion to billing
• SLA monitoring — materialized-view dashboards refreshed by pg_cron, per-project/city/kecamatan scoping
• Billing workflow — Pra-PI, PI with batch, SI, Lunas reconciliation for finance
• AI operations agent — LangGraph/DeepAgents.js with subagents, checkpointer memory, MCP tools, task listener
• Rust analytics core — petgraph knowledge graph (warehouses, cities, ports, airports, pathfinding) + parallel Monte Carlo route simulation (rayon, 10K+ iterations, optional C++ FFI)
• Client view-only portal + external Pusbuk API — read-only JSON:API 1.1 with token rotation, rolling quota, and a synthetic-data sandbox environment
• Driver tracking — GPS telemetry, Leaflet maps, anomaly detection, WhatsApp admin routing by region
• WMS microservice suite — serial registry, AI-assisted verification, incident, inventory, webhook ingestion, hash-chained audit log, BAST/government report services

Technology

• React 18 + Vite + TypeScript — dashboard & field PWA
• Tailwind CSS + Radix UI, TanStack Query, Recharts, Leaflet
• PostgreSQL via Supabase — RLS, PL/pgSQL RPCs, materialized views, realtime, storage, Deno edge functions
• Go (Fiber) — API gateway: JWKS auth, adaptive AIMD rate limiting, SSE, OpenTelemetry; agent-server; WMS services
• Rust — Axum API server, petgraph knowledge graph, Monte Carlo engine (rayon), hash-chain audit service
• C++ — simulation core via FFI
• Bun + Hono — client-facing Pusbuk 2026 API (JSON:API 1.1, Swagger)
• DeepAgents.js + LangGraph + LangChain — AI agent with subagents and MCP tools; OpenRouter LLM, LangSmith tracing
• Python — data pipelines, fleet/dataset normalization & validation scripts
• gRPC / ConnectRPC + Protobuf — abc.agent, abc.graph, abc.montecarlo, antero.tasks contracts
• Docker Compose + Caddy on VPS — production topology; Kubernetes manifests + Terraform (EKS/RDS/VPC/IAM) for scale-out
• Sentry + OpenTelemetry — observability across all services; Vitest, go test, cargo, bun test, GitHub Actions CI

Technical Decisions

• Offline-first over always-online: Couriers deliver to remote kecamatan with unstable connectivity — the PWA treats the network as optional, queueing forms and compressed photos in IndexedDB and syncing on recovery.
• Postgres as the enforcement layer: Client scoping, role guards (is_client(), client_project_keys()), and project isolation live in RLS policies and security-definer RPCs — not just in the API — so external viewers and internal roles share one enforced model.
• Polyglot by workload, not by fashion: Rust owns graph/simulation compute, Go owns the edge (auth, rate limiting, proxying), TypeScript owns UI and agent logic — connected through explicit gRPC contracts.
• AI advises, workflows decide: The agent forecasts routes, queries shipments, and answers operational questions — mutations still flow through auditable RPC and edge-function paths.
• Scale engineered ahead of volume: Trigram indexes, keyset pagination, thumbnail media pipeline, realtime bulk-op guards, and append-only retention were designed against a measured 80K-shipment / ~1 TB media capacity model, not guessed.

Impact

Screenshots