Systems I've shipped.

Context

Executives, managers and analysts wanted answers from production data without waiting on engineering. The first version of the natural-language-to-SQL assistant returned plausible but wrong numbers.

What I did

• ·Built a remote, read-only MCP server inside the FastAPI backend with 8 tools (query, cursor pagination, fuzzy user search, schema discovery), API keys issued from the admin portal, and per-tenant scoping enforced by rewriting every SQL statement with sqlglot.
• ·Traced the wrong answers to root causes, not prompts: a missing business rule caused a 98.4% undercount of approvals, an INNER JOIN dropped records, and the model was offered views that didn't exist. Fixed them with a single source of business rules embedded in the server, guarded by a drift test.
• ·Ran an adversarial security review of my own server: closed scope-bypass and PII-exposure paths and tested 30+ attack shapes with zero leaks.
• ·Built an autonomous morning triage agent (headless Claude Code with the Sentry, Jira and Slack MCPs). It root-causes production errors against the repo and files at most 3 tickets per run through a 4-condition gate, and it is not allowed to commit code or open PRs.

Context

Agents can write code. The hard part is who reviews it. I built the harness around that question: agents plan, build and review each other's work under explicit gates.

What I did

• ·Built an adversarial evaluator panel: two blind judges review every branch independently, a synthesizer confirms findings, and a separate fixer applies only confirmed defects. The loop repeats until zero confirmed CRITICAL findings remain.
• ·Built 20 purpose-built agents, including a simulated C-suite (CTO, CPO, CMO, finance, legal, trust & safety) for product decisions, and synthetic user personas that QA the product end to end.
• ·Built an autonomous backlog runner that pulls issues from Linear and works them in isolated workspaces. First full run: 13 issues → 11 merged PRs in one day.
• ·Wired lifecycle hooks on 10 events, 3-tier model routing, vector memory, and spec-driven development (OpenSpec).
• ·At Prosperas: a swarm harness plus a weekday Sentry triage agent (max 3 tickets per run; it cannot commit or open PRs) and MCP agents over production data.

Context

A major telecom operator wanted to offer credit to the users of its Super App — fully digital, with no paperwork and no branch visits.

What I did

• ·Own the end-to-end AWS architecture with CDK and standardized multi-environment provisioning (dev → staging → prod).
• ·Implemented user authentication with Cognito and secrets with Secrets Manager inside CI/CD, with CloudWatch alerting.
• ·Optimized analytical queries over 2M+ records to cut dashboard latency and AWS cost.
• ·Lead the engineering team and act as the technical interface with the partner.

Context

The AWS bill had grown for years across 17 enabled regions, with no budget, no anomaly alerts and no one who could say what each resource was for.

What I did

• ·Ran a read-only sweep of every enabled region, then removed spend in order of evidence. I retired a BI service that I had already replaced, closed an unused account, deleted a database with zero connections in 13 months, and turned off non-production insights.
• ·Refused a tempting downsize. The production cluster averaged 12.7% CPU, but its p99 was 97.5%.
• ·Put guardrails in place so it doesn't creep back: an anomaly monitor, a monthly budget, log retention on 240 of 240 log groups, and ECR lifecycle rules that took the registry from 229 GB to 21 GB.

Context

A marketplace where AI that actually knows each user fights loneliness. It runs several internal micro-economies, so the product and the business model are built together. Creators design AI personas that hold long-running conversations. One engineer: me. So the architecture, the infrastructure and the delivery process had to make one person as effective as a team.

What I did

• ·Designed a hexagonal modular monolith (Next.js, TypeScript). Its 10+ bounded contexts have dependency rules enforced in CI, and every LLM call goes through a port, so switching models is a config change.
• ·Designed two-tier conversational memory: structured facts plus a rolling summary are always loaded, and semantic recall from pgvector is top-k only, each layer with its own token budget. A per-user encryption key makes a deletion request a key deletion (crypto-shredding).
• ·Built a mandatory safety gate before any LLM call. Blocked messages are never persisted and never reach the model, and every block is audited.
• ·Built payments on an append-only ledger with signed, idempotent webhooks, verified adversarially: replays, forged signatures and tampered amounts all fail safely.
• ·Built provider-portable infrastructure (Docker, Caddy, Cloudflare, Terraform). The app moved across three hosting providers without a rewrite, and rollback is exercised in production.
• ·Run delivery with a multi-agent software factory: planner, coder and adversarial reviewer agents working from specs. In the first full run, 13 issues became 11 PRs merged the same day.