Skip to content

v0.1 — Discovery (SUPERSEDED by v1.0)

Captured from the grilling sessions of 2026-11-11 → 2026-11-12. Kept for history: shows how the design moved.

Brief

Production-grade MCP server on AWS with proper auth and system-design patterns (timeouts, circuit breaker, retry…), realistic fintech data (transactions, KYC), and a RAG knowledge base over department SOPs. Public GitHub showcase for recruiters: aesthetic, technically deep, rich infographics and animations.

Round 1 — accepted

  • Guidance level HIGH · Region eu-north-1 (project carpe-diem) · idle budget target < $10/month (later revised)
  • Fictional Indian neobank (UPI/IMPS/NEFT, PAN/Aadhaar, RBI KYC MD, PMLA)
  • MCP Authorization spec with Cognito, per-tool scopes, role-based PII masking, audit log
  • Showcase: README, docs site, demo GIF, load/chaos report, live endpoint
  • Versioned plans in docs/plans/, ADRs, glossary, MIT, GitHub Actions + OIDC

Changes along the way

Initially proposed Changed to Why
TypeScript end to end Python (FastMCP + FastAPI) Owner's expertise; must defend every line in interviews
DynamoDB single-table + DuckDB/Parquet PostgreSQL system of record pgvector needs Postgres anyway; ledger wants ACID, joins, aggregates
LanceDB on S3 / S3 Vectors pgvector S3 Vectors unavailable on plan; one datastore
Titan Text Embeddings v2 OpenAI text-embedding-3-large Owner preference; quality
No queue SQS + DLQ for ingestion and outbox events Real async resilience story; Kafka rejected
k6 Locust Keep one language
Astro Starlight MkDocs Material Keep one language
< $10/month idle ≈ $25/month + pause/resume RDS + NAT instance required by pgvector-in-VPC
Simple transaction statuses Rich lifecycles (on_hold, credit_delayed, deemed_success, …) Owner request; realism