Case study
Automation & Agent Portfolio
Around thirty low-code automations, six voice agents and two resilient scraping pipelines wiring LLMs into real business workflows.
- Role
- Sole builder
- Domain
- Automation & Agents · Generative AI & LLMs
- Stack
- n8n · Retell AI · ElevenLabs · Google Gemini · OpenAI
- Result
- ~30 automation workflows
Overview
A deliberately broad collection of automations and AI agents, built solo to learn which tool fits which problem — from low-code workflows to voice agents to Python where low-code runs out.
What is in it
Lead generation & outreach
Find prospects, enrich them, write to them.
- Search and profile data pulled, deduplicated, then structured by an LLM
- Personalised outreach drafted per prospect
- Everything lands in a sheet a person can review
Content & publishing
From a source or an idea to posts on several platforms.
- Scheduled post generation with structured output
- Articles rewritten for social formats
- A multi-platform video publishing chain with error branches
Conversational agents
Agents that answer over documents and tools.
- Chat with a PDF — split, embedded, retrieved, answered
- Messaging agents with routing and short-term memory
- An agent calling an external scraping tool over a tool protocol
Voice & operations
Phone agents, orders and invoices.
- Six voice agents with conversation flows, keypad input and call analysis
- Store orders pushed into a database and email
- Invoice creation in an accounting system from a database trigger
Patterns that repeat
- Structured output
- Every LLM step returns a parsed schema, never free text the next step has to guess at.
- Rate limits
- Batching and deliberate waits wherever an external API has a quota.
- Dedupe before write
- Nothing is written twice, however many times a run repeats.
- Explicit failure
- Error branches stop a run loudly instead of letting bad data flow on.
My Role
Role
Sole builder
Contribution
- Built triggered, LLM-augmented pipelines across lead generation, multi-platform content publishing and order/finance automation
- Built a RAG chatbot over an embedding store with retrieval-QA
- Built an agent that calls an external scraping tool through a tool-use protocol
- Configured six conversation-flow voice agents with DTMF, multi-language and post-call analysis
- Built two resilient scrapers with stealth, CAPTCHA handling, resumable checkpoints and a scrape-then-enrich stage
Team Context
Sole builder. Client-specific workflows within the collection remain confidential and are not shown.
What the breadth taught
- Low-code is right until state gets involved. The two scrapers needed resumable checkpoints across long runs, which is where Python took over.
- Retrieval is only worth it when the source is too big for the prompt. Several early ideas were simpler as one well-shaped prompt.
- Voice agents need designed conversation flows, not open chat. Callers go off-script in predictable ways.
Results
~30
automation workflows
6
voice agents
15+
SaaS integrations
Limits
A showcase collection rather than one production system — most workflows were built to demonstrate a pattern and were never measured in sustained use. There are no automated tests, and the scrapers depend on markup the target sites can change at any time.
What I'd Improve
- Keep workflows in version control as code, not exported snapshots. First fix.
- Share one error-handling sub-workflow instead of re-implementing it in each.
- Evaluate the document chat answers against a small question set, rather than judging them by eye.