Bob's Lifebot — High-Level Design

Architecture

flowchart TB
  subgraph Sources
    HEVY[Hevy API] ; HAE[Health Auto Export iOS] ; STUDY[Udemy userscript]
  end
  subgraph Browser
    SPA[Web SPA<br/>CloudFront + S3]
  end
  GW["API Gateway (HTTP API)<br/>throttled 20/40"]
  subgraph Lambdas["Lambdas · python3.12 · arm64 · perms-boundary"]
    ING[ingest] ; ADV[advice] ; CHAT[chat] ; STATS[stats] ; HSYNC[hevy_sync]
  end
  DDB[("DynamoDB<br/>lifebot-records")]
  S3[("S3 raw archive<br/>versioned · AES256")]
  BED["Bedrock<br/>Claude Haiku 4.5"]
  SSM["SSM SecureString<br/>/lifebot/*"]
  COG[Cognito user pool<br/>PKCE · admin-create-only]
  CRON[EventBridge daily]

  HEVY & HAE & STUDY -- "x-lifebot-key" --> GW
  SPA -- "Cognito JWT" --> GW
  SPA -. login .-> COG
  GW --> ING & ADV & CHAT & STATS
  CRON --> ADV
  HSYNC --> DDB
  ING --> S3 & DDB
  ADV & CHAT --> BED
  ADV & CHAT & STATS --> DDB
  SSM -.-> ING & ADV & HSYNC

Auth model (two tiers)

1. Machine ingest — POST /ingest/{health,hevy,study} and POST /advice are gated by a shared x-lifebot-key header (an SSM SecureString), because the callers (Hevy, Health Auto Export, a Udemy userscript) are machines with changing IPs — a shared secret is the right M2M auth, not IP-locking. Ingest is idempotent (de-dupes re-sends).
2. Browser app — GET /api/stats, POST /api/goals, POST /api/chat require a Cognito JWT (API Gateway JWT authorizer). The user pool is admin-create-only (no public signup), single user, 12-char password policy, PKCE public SPA client, short tokens (id/access 1 h, refresh 30 d). Login via the Cognito hosted UI.

Components

• API Gateway HTTP API (cheaper than REST) — CORS restricted to the app domain, the CloudFront domain, and www.udemy.com (the study userscript); throttled (rate 20 / burst 40) to cap Bedrock + DynamoDB cost/DoS.
• ingest — archives raw JSON to versioned S3, normalizes to daily records in DynamoDB.
• advice — daily digest (EventBridge) + on-demand POST /advice; Bedrock Haiku; stores advice.
• chat — POST /api/chat, Q&A over your recent records via Bedrock (lookback window).
• stats — GET /api/stats (dashboard) + POST /api/goals (branches on HTTP method).
• hevy_sync — pulls Hevy workouts into the store.
• Storage — DynamoDB lifebot-records (on-demand, 25 GB free tier); S3 raw archive (versioned, AES256, private).
• Frontend — CloudFront + S3 SPA; runtime config.js templated by Terraform.

Security

• Permissions boundary on every Lambda role — a ceiling the deploy user itself cannot exceed. Each role also gets a least-privilege inline policy scoped to the one DynamoDB table ARN and only the actions it needs (ingest write-only; chat/stats read; etc.).
• Bedrock scoped to the Haiku model/inference-profile ARNs only; KMS Decrypt gated by a kms:ViaService = ssm condition (can only decrypt SSM params).
• Secrets in SSM SecureString, never in Terraform state; Cognito password set out-of-band.
• S3 private + versioned + AES256; raw health payloads never logged anywhere world-readable.
• AI privacy — Bedrock/Anthropic do not train on API data (confidential by default), which is why Bedrock Haiku was chosen over an unpaid Gemini tier for health data.
• Full findings + minor items in Bobs-lifebot/security-review.

Cost

Target ≤ £5/mo, typically pennies: HTTP API + Lambda (free tier covers personal volume), DynamoDB on-demand (25 GB free), S3 (pennies), SSM standard (free), Bedrock Haiku (~pennies at ~1 call/day on aggregated data). No always-on DB (RDS/Aurora was explicitly rejected as budget-blowing). Deliberately aggregate to daily/hourly before any AI call — never dump raw HealthKit samples into a prompt.

Related