AI-driven automation systems built with Claude, n8n, and self-hosted infrastructure — each shown as problem, approach, and outcome.
Credit Analyst Meeting Handover Automation
Live-tested end-to-end with a real meetinggithub.com/anthchilaka/credit-analyst-meeting-handover-automationProblem
Credit analysts received manual write-ups after client calls they hadn't personally attended, producing inconsistent, incomplete handover records with no standard structure.
Approach
Built an event-driven pipeline — a Fathom call-recording webhook triggers n8n the moment a call ends, which runs a structured summary (overview, financial/credit details, decisions, action items, open questions) and emails it out with zero manual effort. Event-driven was chosen deliberately over polling for lower latency and lower running cost. Runs on Docker Compose (Postgres, Redis, n8n queue mode) on a self-hosted Hetzner VPS with Nginx SSL.
Outcome
Processes a call within under a minute of it ending, at an infrastructure cost of about $7.09/month and a few cents of Claude API usage per meeting. In testing it correctly reported "no financial/credit-relevant details discussed" on a call where that was true, rather than inventing content — the trait that matters most in a document analysts will actually rely on. Validation so far is one real meeting plus schema-matched synthetic payloads; CRM integration and wider meeting-length coverage are explicitly out of scope for now.
Stack: n8n, Claude API (Sonnet 5), Fathom webhook, Gmail API, Docker Compose, PostgreSQL, Redis
Claude Workspace Audit
Methodology validated, production run 2026-07-09github.com/anthchilaka/claude-workspace-auditProblem
Claude Cowork/Code builds accumulate operational drift — unrotated scratch folders, duplicated skills, broken paths, missing test coverage — with no repeatable way to catch it, and any audit tool has to run under a strict propose-only constraint so it never acts without explicit approval.
Approach
Built a Claude skill implementing a 6-point build-hygiene checklist that runs entirely on the user's own machine, requiring no external file access. Evaluated three ways: synthetic fixture testing for detection accuracy, real production validation against live daily job data, and cross-vendor grading gates spanning Claude, OpenAI, and Google models. A baseline comparison found the entire gap traced to propose-only discipline, where the baseline deleted an entire target folder without asking, overriding a delete-permission prompt. A production validation run on 2026-07-09 against the linkedin-job-intel skill found and fixed three real rule gaps, including a keyword match where 11 of 12 hits were false positives.
Outcome
Cross-vendor grading gates spanning Claude, OpenAI, and Google confirm the propose-only discipline holds under adversarial testing, not just in the happy path.
Stack: Claude Code / Cowork, cross-vendor model grading (Claude, OpenAI, Google), synthetic fixture testing, production data validation
Kachi AI Assistant
Live on Telegram · WhatsApp in progressgithub.com/anthchilaka/kachi-ai-assistantProblem
CraftByTag, a Nigerian retail business, needed to meet customers on the messaging apps they already use rather than routing them to a website, and to personalize returning-customer interactions without a staff member manually triaging every inbound message.
Approach
Built a conversational AI assistant on a queue-mode n8n instance (Redis-backed job broker) that runs a real Claude reasoning call to validate each customer's free-text description before capturing structured data, verifying email, and confirming next steps — no human in the loop. Handles three intake categories (AI automation guidance, craft training, freelance services) with voice input transcribed via OpenAI Whisper.
Outcome
Live and operational on Telegram at @ask_kachi_bot — the AI-automation intake flow is confirmed working through real customer conversations, not isolated testing. WhatsApp is in progress pending Meta Business Portfolio verification, launched on a deliberate staggered-rollout call rather than holding both channels for a simultaneous release.
Stack: n8n, Redis, PostgreSQL, Anthropic Claude, OpenAI Whisper, Telegram Bot API, WhatsApp Cloud API
