Oluwaseyi (Ekpere) Oyewunmi

Software Engineer · data platform & DevOps · Aktivate · Dover, DE · Jan 2025 — Sep 2026

An influencer-analytics platform for African-market brands. I built its data engine on my own: a FastAPI service that scrapes Instagram, TikTok, X, YouTube and LinkedIn on a self-protecting schedule and turns raw posts into engagement, authenticity and topic scores. Then I moved the product's AdonisJS and Nuxt apps from PM2 to health-gated container deploys.

• ▹Sole author of the scraping engine: 5 platforms behind a 20-endpoint FastAPI service, with MongoDB upserts deduplicated on (platform, post_id)
• ▹Anti-bot strategy per platform: Instagram's private GraphQL through Scrapfly geo-proxies, TikTok via a JS auto-scroll that captures the page's own item_list XHRs, X via Playwright with masked automation flags and cookie-session reuse
• ▹Self-protecting scheduler: 20-job batch cap, a circuit breaker that pauses when more than 20% of the last 50 runs fail, 2ⁿ-hour backoff, and a per-platform X concurrency cap tuned down in production
• ▹Transparent scoring: engagement rate plus a consistency score from the coefficient of variation, a weighted authenticity score, and hybrid topic tagging (keyword rules + Mistral-7B, falling back to rules if the model fails)
• ▹Moved the product off PM2: GHCR images tagged by commit with layer caching, health-gated Compose deploys isolating API, queue worker and two Redis instances. Trialled Docker Swarm first and rolled back within 9 days

Co-lead Engineer · SwarmTech · Remote · Aug 2024 — May 2025

Bot Raiding as a Service. A crypto project enrolls its Telegram group, a moderator points the bot at a post on X, and it watches engagement until the targets are hit, grades every raider's reply, and maps the scores to token payouts on Solana. A two-person codebase: I owned scraping, AI grading, deployment and on-chain funding.

• ▹11 per-chat, per-user conversation flows, including an 8-step enrollment state machine; bit-flag roles checked against Firestore, then version and subscription gates, before every privileged command
• ▹Real multi-tenancy: the tenant is resolved from the Telegram chat, with its own raid targets, time limits, scoring weights and payout bands, and separate DEV/UAT/PROD bots and Firebase projects
• ▹Raids run as cancellable background tasks: a stealth Playwright scraper (masked automation flags, cookie-session reuse, human-like scrolling) polls every 30s until replies, likes, bookmarks and reposts all hit target, and even opens X's hidden "probable spam" replies so they're excluded
• ▹Cost-gated hybrid grader: tenant-configurable rules score every reply, a local Llama 3.2 (Ollama) only sees the low scorers, and the 60/40 blend maps to 7 payout bands, degrading to rules-only if the model fails
• ▹On-chain funding via Helius webhooks sharded up to 100K watched wallets each, re-verifying balances before automatically activating a tenant's plan

Lead Backend & DevOps Engineer · Hordanso · Remote · Oct 2023 — Oct 2024

A self-serve platform for reselling Google Workspace. Customers check a domain and pay through Stripe, and a webhook provisions everything on Google's side (customer, super admin, subscription), then keeps seats in sync. I led the FastAPI backend and built the Kubernetes infrastructure and CI/CD the platform's microfrontends run on.

• ▹Payment-driven provisioning: a signature-verified Stripe webhook creates the Workspace customer, promotes a super admin and opens the subscription, then writes IDs and billing dates back to Firestore
• ▹Integrated 3 Google APIs (Reseller, Admin Directory, Site Verification) through a domain-wide-delegated service account: 3 Workspace editions × 3 billing plans, seat and plan changes, suspend and reactivate
• ▹Stripe quantity changes sync to Google seats, and a Celery beat job on Redis flags annual commitments ending within 7 days and trials within 20
• ▹One domain check that combines registrar availability with "is this already on Workspace?"; Firebase auth with TOTP 2FA and role-guarded routes
• ▹Hetzner k3s cluster across 3 locations with an autoscaling worker pool (0–5 nodes), ingress-nginx behind a load balancer, and cert-manager TLS for 8 microfrontends, each with its own path-filtered build-and-deploy pipeline

Lead Developer · RouletteAid · Madrid, ES · Jul 2022 — Feb 2024

A Windows desktop app that predicts the next spin per bet type from the last three spins, keeps learning from every prediction, and ships in three paid tiers with time-metered licences and a Git-based auto-updater.

• ▹One model per bet family, 10 in all: k-NN, an SVM pipeline, LightGBM, CatBoost and Random Forests, each chosen for its own bet type
• ▹21 wheel-geometry features from the last 3 spins: forward and backward pocket distance on the 37-pocket European wheel, neighbours, figure and subset
• ▹Online learning: each prediction is appended to its model's 5,000-row seed set and the model refit; the top tier works on copies restored at exit, so a session can't corrupt the shipped models
• ▹Tiered, time-metered licensing: 3 tiers × 4 durations that stack, validated against the licence server with a network clock to resist local clock tampering
• ▹Self-updating installs: change detection against a private repo, a shallow single-branch clone on a background thread with progress, packaged with PyInstaller bytecode encryption into a Windows installer

Lead Software Engineer · FeteFlex · Lagos · Jun 2024 — Nov 2025

• ▹API response times −70% while holding 99.9% uptime
• ▹Docker + Kubernetes migration cut AWS costs ~80% and deploy time ~40%

Software Engineer · GreatSomme Inc · Lagos · Mar 2021 — Nov 2022

• ▹Proxy personas + relational profiles → a tailored dashboard per role
• ▹E-learning module with a 500+ book class library, full-text search, analytics across 20+ metrics