ebube
← Back to ExperienceRole 01 · 2026 - Present
Remote · 2026 - Present

Yumb

Founder & Fullstack Engineer

Solo-built AI cold outreach SaaS that automates lead discovery, prospect matching, and multi-step email campaigns. The backend is written in Go 1.25 with Chi and pgx connection pooling. To handle unreliable LLM providers, I designed a 4-tier fallback architecture cascading from Anthropic to Groq to OpenRouter, with an offline template generator as a final safety net. Outbound messages send directly via Gmail OAuth2 with thread continuity and rate limiting.

4 Tiers
LLM cascading failover architecture

Role Overview

Yumb is an automated cold outreach platform handling lead scraping, resume-to-stack TF-IDF matching, personalized multi-channel copy generation, and Gmail sequence delivery. Currently running in production with Paystack billing in test mode.

Architecture & Implementation

The backend runs Go 1.25 with the go-chi router and PostgreSQL 16 managed by jackc/pgx pools. An automated engine waterfalls requests through Anthropic, Groq, OpenRouter, and a deterministic template generator when external APIs fail. Concurrency is split across two dedicated worker pools: one for initial drafts (concurrency 3) and another for follow-up sequences (concurrency 4, capped at 25 items per batch). Authentication supports Google OAuth2, LinkedIn OIDC, FIDO2/WebAuthn passkeys, and Argon2id passwords with email OTP. Sentry captures telemetry across both services, and GitHub Actions runs CI with live PostgreSQL container testing and race detection.

Frontend & Mobile Integration

Draft Studio and the sequence management console are built in React 19 and TypeScript, using Zustand for predictable local state and fast prospect filtering.

Key Outcomes

  • ▪Cascading LLM engine: automatic failover across Anthropic, Groq, OpenRouter, and an offline heuristic generator so drafting never halts during API outages.
  • ▪Worker pools handle batch generation with concurrency limits (3 workers for initial drafts, 4 for follow-up sequences, max 25 items per batch).
  • ▪Direct Gmail OAuth2 integration with in-thread replies and account-tiered rate limits (40/hour, 500 to 2,000/day).
  • ▪Automated CI pipeline in GitHub Actions running go vet, live PostgreSQL container tests with race detection, and frontend verification.
← Previous roleRoslidochNext role →Revela Africa