Shirisha Gujja

Problem

Developers switch between multiple AI coding tools — Claude Desktop for one task, Cursor for another — and each one starts from zero, because nothing carries context between them. A decision explained to one tool is invisible to the next, and re-explaining it every session defeats the point of an assistant having memory at all.

What I built

• A PostgreSQL-backed memory service exposed over MCP, so Claude Desktop, Cursor, and any other MCP client read and write one shared, versioned project memory instead of each maintaining its own.
• Optimistic concurrency on every write: a compare-and-set update means two clients editing the same memory at once produce one winner and one explicit conflict response with the winning revision attached, never a silent overwrite.
• Immutable revision history with structural supersession — retiring a memory excludes it from every retrieval path in the same transaction that creates its replacement, so an outdated decision can't resurface as a search result while remaining fully readable in the audit trail.
• Hybrid retrieval combining PostgreSQL full-text search and pgvector similarity search, merged with Reciprocal Rank Fusion and served through a token-budgeted context tool so results fit a caller's context window.
• Idempotent writes and content-hash deduplication kept as two distinct mechanisms: one client retrying the same request replays the original response, while two clients independently asserting the same fact are merged as corroborating evidence rather than duplicated.
• 369 tests against a real PostgreSQL instance, including 50 concurrent writers proving the conflict resolution for real and a hand-graded retrieval-quality benchmark gated against a committed baseline.

Architecture

Problem

Personal finance data is scattered across banks, brokerages, and manual tracking. Getting one accurate picture — spending by category, budget adherence, net worth, investment performance — usually means a spreadsheet, or handing a third party read access to every account.

What I built

• A modular backend architecture covering accounts, transactions, budgets, savings goals, investments, net worth, and cash-flow forecasting, all scoped to the authenticated user.
• Plaid integration for bank account linking and transaction sync, with encrypted access tokens, cursor-based pagination, and database-level deduplication so a re-synced account never creates duplicate transactions.
• A transactional outbox that writes each event in the same database transaction as the record that triggered it, so no event is lost if the message broker is unavailable.
• Three independently deployable consumer services on Kafka that categorize transactions, detect unusual spending with idempotent alert creation, and push real-time alerts to the browser over WebSockets.
• AI-generated monthly spending insights grounded on pre-computed aggregates, with a deterministic template fallback when the model call fails.
• Distributed tracing, metrics, and structured logs correlated end to end through OpenTelemetry, Prometheus, Grafana, and Tempo, with failure testing that validates recovery from crashes, outages, and broker failures.

Architecture