← Artifact library

Technical design / Proposed stack

Basketball AI coach

Youth basketball video → computer vision → coaching insights.

Proposed architecture. Camera hardware, CV model and assistant provider remain open.

Technical map

Proposed target states

Systems are grouped by layer. Arrows show the integration method and the data exchanged.

Detailed engineering view. Logical components share servers; each box does not imply a separate server.

Tech stack

Proposed

Phases describe product maturity, not fixed dates. Existing components carry forward; later stages add features and capacity.

Cost estimate

USD · Guesstimate

Budget reference, not a market-average survey or supplier quote. Prices checked 27 October 2026. Adjust the scenario below; all stages are proposed configurations.

Assumptions, hardware & exclusions

Server placement: CPU budget combines Next.js, API, ingest, Redis/RQ and MLflow/CVAT on self-managed VMs. Mid uses two $48 VMs; no separate managed Redis fee is included. Supabase covers app metadata and login. Training and inference use rented GPU capacity, not an owned GPU server.

GPU: RTX 4090 planning anchor $0.89/hour. GPUs must be stopped when idle. 730 GPU-hours cost $650 for one always-on GPU before disk and transfer. Mid assumes 4 inference GPU-hours/game plus 80 training hours/month; this is not a measured throughput claim. Larger models or retries increase the total.

Storage: 90-day steady-state retention, 25% derived-media overhead and the scenario’s fixed research dataset. First-year TCO uses this steady-state run rate for all 12 months, conservatively ignoring initial storage ramp-up. $0.025/GB-month is a rounded S3 planning allowance, not a confirmed regional quote; $0.09/GB egress is an AWS public pricing reference. Includes one full footage read for inference and one playback equivalent, plus training reads. Free allowances are not credited.

Hardware CAPEX per venue: camera $300 + capture device $250 + local storage $150 + mount/network/power hardware $100 + installation $200 = $1,000. Rough allowance range $500–$1,800, not a selected hardware quote. Assumes one camera and existing internet; court coverage, recorder compatibility, installation and local prices need validation. Existing hardware can reduce incremental CAPEX.

Excluded: taxes/VAT, import/shipping, exchange-rate changes, footage licensing, annotation/review labour, engineering, support labour and enterprise availability. These totals are infrastructure + capture-hardware TCO, not full project cost. Hardware CAPEX is cash spend; accounting treatment and depreciation are not modelled.

Optional people-cost illustration: initial engineering 240 hours × $40 = $9,600 one-time; pilot annotation/review 40 hours × $20 + operations 8 hours × $40 = $1,120/month. These are editable planning assumptions, not market salary rates. Add $23,040 to a 12-month pilot TCO if these assumptions apply. Engineering spend is not automatically classified as CAPEX.

Phased first year: per-scenario TCO means 12 months at that stage; do not add three 12-month totals. A staged plan can use 3 × short monthly OPEX + 6 × mid + 3 × long + hardware purchased once. People costs remain separate.

Price references: DigitalOcean CPU VMs · Runpod GPU & disks · Supabase · AWS S3. Placement across providers introduces egress and operations overhead.