# Engagement Models & Pricing Guide — AstrumFlow (广西星域跃迁智能科技有限公司)

Last updated: 2026-11-19
Contact: contact@astrumflow.com | https://www.astrumflow.com/en/contact | https://www.astrumflow.com/contact

AstrumFlow delivers custom full-stack systems across AI/LLMs, software platforms, embedded IoT hardware, and cloud infrastructure. Because enterprise and hardware-software systems vary in scope, engagements follow a transparent 4-stage milestone structure with clear deliverables and 100% IP handover.

---

## 1. Engagement Tiers & Milestone Structure

### Tier 1: Technical Diagnosis & Architecture Blueprint
- **Price**: Free initial consultation & feasibility assessment
- **Turnaround**: 3–5 business days
- **Included**:
  - Direct consultation with lead AI, software, or embedded hardware architects (no non-technical sales gatekeepers)
  - Mutual NDA signing prior to reviewing proprietary data, workflows, or schematics
  - On-site or remote workflow mapping and bottleneck ROI analysis
  - Written architecture blueprint, deployment topology (on-premises / cloud / edge), and milestone quote

### Tier 2: Real-Data MVP / Proof-of-Concept (PoC) Sprint
- **Typical Timeline**: 2–4 weeks
- **Billing Model**: Fixed-scope milestone fee (credited toward full build upon expansion)
- **Target Use Cases**:
  - Enterprise RAG knowledge base or Feishu/WeCom agent tested on 100–1,000 internal documents
  - Single-line Edge AI visual defect inspection model trained on 30–100 defect samples
  - Custom digital-human anchor cloning (1080p avatar + TTS voice model) and live-room test
  - Embedded sensor node & PCB prototype schematic / breadboard validation
- **Deliverables**:
  - Working prototype running on your real operational data
  - Quantitative benchmark report (accuracy, latency, recall, or hardware power profile)
  - Go / No-Go production sizing and BOM/compute cost calculation

### Tier 3: Turnkey Full-Stack Engineering Build
- **Typical Timeline**: 4–12 weeks (depending on software/hardware complexity)
- **Billing Model**: Milestone-based (e.g., Architecture Sign-off → Alpha/EVT Prototype → Production Deployment → Acceptance)
- **Deployment Options**:
  - **Air-Gapped On-Premises**: Deployed inside client server rooms, hospital intranets, or private VPCs
  - **Public / Hybrid Cloud**: Containerized Kubernetes deployment on Alibaba Cloud, Huawei Cloud, Tencent Cloud, AWS, or Azure
  - **On-Device / Edge**: Quantized INT8/INT4 models on industrial edge boxes or FreeRTOS/Zephyr firmware on STM32/ESP32/nRF MCUs
- **Full IP & Asset Handover (Zero Vendor Lock-in)**:
  - 100% of custom source code, database schemas, and OpenAPI specifications
  - Fine-tuned model adapters/weights, RAG indexing pipelines, and prompt/guardrail configurations
  - Complete hardware design files: schematics, multilayer PCB layouts (Gerber), pick-and-place files, and Bill of Materials (BOM)

### Tier 4: Managed Operations, AIOps & Continuous Iteration
- **Billing Model**: Annual SLA retainer or on-demand engineering support
- **Service Targets**:
  - 99.95% high-availability monitoring via Prometheus, Grafana, and OpenTelemetry
  - Automated backups and disaster recovery drills (RPO < 5 minutes, RTO < 15 minutes)
  - Continuous model drift evaluation, RAG knowledge base synchronization, and seasonal hardware/firmware OTA updates

---

## 2. Reference Compute & Hardware Sizing Benchmarks

| System Type | Typical Deployment Footprint | Key Performance Benchmark |
|---|---|---|
| **Enterprise RAG & AI Assistant** | 1–2× Enterprise GPUs (on-prem) or Private VPC API | >92% Top-3 citation recall; <1.5 s first-token latency; 100% ACL permission inheritance |
| **7B–72B Domain Fine-Tuned LLM** | INT8 / INT4 Quantized (vLLM / TensorRT-LLM) | 50%–75% VRAM reduction vs. FP16; >98% task accuracy retained |
| **Edge AOI Visual Inspection** | Industrial Edge Box + Global-Shutter Camera | <15 ms per-frame inference; ≥99.4% defect recall; <0.5% false alarm rate |
| **Low-Power Field IoT Node** | Custom STM32/nRF PCB + Solar + LoRaWAN | <5 μA deep-sleep draw; 3–5 km gateway range; >30 days rainy-weather battery reserve |
| **1080p Digital-Human Studio** | Real-Time Lip-Sync & TTS Cloning Pipeline | 3–5 min training footage; <120 ms audio-visual sync latency; 24/7 live streaming |
