Systems

Engineering teams lose hours to interrupts: Sentry alerts nobody triages, questions that pull senior engineers out of focus, pull requests that wait days for review. I built an agent that absorbs much of that load, in daily production use across Slack, Sentry, and GitHub. It triages errors as they fire, answers engineering questions in Slack threads with real codebase context, and reviews pull requests.

Most of the design work went into making the autonomy trustworthy. The fix pipeline spawns coding agents (Claude Code or Codex CLI) in isolated git worktrees so nothing touches a shared checkout. A multi-pass self-review gate runs before a human sees anything, and the bot opens draft PRs that it reworks itself when reviewers request changes. The tool architecture is brain-agnostic: LLM backends sit behind one interface and can be swapped without a rewrite.

Nothing crashes when an agent system overspends. The invoice just arrives, and it's wrong. I designed the telemetry layer that catches this early. The core decision was to store raw token counts instead of computed costs, so a provider price change needs no migration and no backfill, and historical data stays accurate. Pricing is applied when the data is read.

On top of that sit anomaly alerts with Redis-backed cooldowns, so a spike pages once instead of every minute, and per-source investigation hints: when spend jumps, the alert names the configuration most likely responsible.

This thinking is also the seed of Caspian, a spend-control and audit product I'm generalizing from it. In development, currently parked while production work takes priority.

Ad platforms optimize toward the conversions you feed them, so the pipeline that moves a CRM deal into an ad platform directly affects revenue. I built Glitch's HubSpot integration end to end: OAuth, encrypted token storage, HMAC-validated webhooks, and conversion delivery into Google Ads. It sits behind a CRM abstraction designed so that adding Salesforce or Pipedrive is configuration work, not a rewrite.

For the Meta Ads expansion I refactored conversion tracking to a polymorphic design that decouples ad platforms from CRM event sources. Any CRM event can feed any ad platform without a separate integration for every pair.

Generating 160+ SEO-optimized topics per client domain, across an enterprise client base, is a queueing problem more than a prompting problem. I built the content-generation systems and the Celery/Redis infrastructure that ran them, with retries, failure recovery, and throughput tuning so a slow LLM call never stalled a client's pipeline. Underneath it, a zero-downtime Django migration strategy let the schema evolve while production data kept flowing.

Led backend development of a platform serving thousands of users. Cut API response times 40% through indexing and query optimization, shipped role-based access control and secure payment integrations, and steered the architecture toward services as the team grew. Also mentored three junior engineers, mostly through code review.

Rebuilt the backend of a payment-acceleration platform (invoicing, collections, financing) in Django while it kept serving businesses. Integrated Spendjuice, Mono, and Paystack, and cut deploy rollout time 30%.

A mobile-first marketplace for live-event photography. People at concerts, games, and festivals capture photos and videos through an in-app camera and publish them to event-tagged feeds. Attendees who missed the shot pay credits to download them, and creators earn real money from their downloads. The social layer drives discovery; downloads are the transaction.

The backend is Django and DRF with PostGIS event matching and Cloudflare Stream for video. Payments run on two rails: in-app purchases through Google Play Billing and Apple StoreKit, verified server-side, with Stripe on the web and Stripe Connect Express handling creator payouts. The mobile client is Flutter, on iOS and Android, and the app is in open testing on Android now.

An LLM-powered content strategy platform. Give it a domain and it builds the whole strategy: business profiling, an entity knowledge graph of 300 to 400 typed nodes, topic clustering, and 50 to 80 content briefs organized as hub-and-spoke clusters with automated internal linking and provenance tracking.

The bet is that topical authority comes from structure. Deriving the content architecture from a knowledge graph instead of keyword lists produces coherent clusters that search engines, and increasingly LLMs, read as depth. It takes the content-systems work I did at Search Atlas and productizes my own version of it.

A ride-sharing platform for the Nigerian market that connects car owners with passengers heading the same direction. The backend is built and tested: Django and DRF with PostGIS geospatial trip matching, wallet payments on Paystack, driver verification, and Celery for async work, plus a production deployment pipeline (Docker, Nginx, GitHub Actions CI/CD). The mobile app is in development; the platform ships when it lands. Ask and I'll show you around.