AI Automation & Voice Agent Engineer

Automation systems, and the software they run on.

Most automation people cannot build the application. Most developers do not do automation. I ship both: n8n and voice systems on one side, Django and Next.js applications on the other, and the systems below are the intersection.

17
automation systems: 8/17
with explicit error handling: ~180k
lines in the OREE product: 100%
Upwork job success

Selected work

13 of 13

VoiceLive in production

CallerHi, can I move my Thursday appointment?
2,109 ms
AgentSure - what day works instead?
CallerFriday morning if you have it.
1,984 ms
AgentFriday has 6 slots before 12:30. Earliest is 9:00.

Voice booking agent

Answers the phone, books into a real calendar, and hands the outcome to n8n over a signed webhook.

median turn: 2,109 ms
offline tests: 201
live intents: 6

FastAPI on Modal
Neon Postgres
Claude Haiku 4.5
Deepgram Flux
Google Calendar
HMAC-SHA256

Turn latency is the measured median on clean audio; the model accounts for 554 ms of it. The exchange above illustrates the reschedule path.

ProductLive

oreeai.com

OREE, the AI SDR platform I built the backend for. Scoring, drafting and the human approval gate in one frame.

OREE

An AI sales co-pilot I built the backend for. It started as 7 n8n workflows, and I rebuilt it as a product when the canvas stopped being the right tool.

lines of code: ~180,000
Django models: 56+
nodes in the n8n original: 228

Django + DRF
PostgreSQL + pgvector
Celery + Redis
Gmail API
Next.js x3
Docker

oreeai.com

ProductPrivate - client work

private - client work

A per-tenant operations view: posts today, posts this month, total posts, pipeline active, a seven-slot daily content queue with times and typed categories, and a per-tenant configuration panel.

The per-tenant operations view. Client identifiers are cropped or redacted.

Content pipeline and operations dashboard

An n8n publishing pipeline and the Next.js screen the client runs it from. One engagement, both halves.

posts published: 456
per month: 180
slots per day: 7

n8n
Next.js
FastAPI
Meta Graph API
Multi-tenant

AutomationPrivate

An n8n canvas: a WhatsApp trigger feeding an intents router, which fans out into colour-coded book, reschedule/cancel, get-message and general branches, each with calendar and sheets tools, plus a scheduled reminder flow.

Error -> Stop and Error

WhatsApp booking agent

A WhatsApp trigger into an intent router, four booking branches, and an error path that stops rather than passing bad data on.

nodes: 53
intent branches: 4
systems with error handling: 8/17

n8n
WhatsApp
Google Calendar
Google Sheets
OpenAI

Opened from exported workflow files in my own instance - credentials do not travel with an export. The warning icons are missing keys, not broken nodes.

AutomationPrivate

An n8n canvas titled OREE Lead pipeline, Bucket 1, Ingestion and Enrichment, Dedupe First. Four parallel lanes each run a client-limit check, a monthly lead count, a daily limit gate and a counter increment, then call a dedupe service and an enrichment step, before converging on a single spine that validates, upserts and calls the research and content generation workflows.

OREE lead ingestion

Four lead sources running as parallel lanes, each gated by a per-tenant quota and deduplicated before a penny of enrichment is spent.

nodes across 7 workflows: 228
lead sources, parallel: 4
dedupe service, shared: 1

n8n
Apollo
PostgreSQL
Webhooks
Multi-tenant

Opened from exported workflow files in my own instance - credentials do not travel with an export. The warning icons are missing keys, not broken nodes.

AutomationPrivate

An n8n canvas for a RAG file ingestion pipeline: a webhook receives an upload, validates it, downloads the file, routes by type between PDF and DOCX, extracts text, then chunks at 1000 characters with 200 overlap, creates OpenAI embeddings and stores them in PGVector. A separate purple group handles errors, logging download and extraction failures and marking the document failed.

Error handling - failed documents marked in the database

RAG document ingestion

Tenant documents into a vector store: routed by file type, chunked, embedded, and with a failure path that records what broke.

chunk size / overlap: 1000 / 200
with Cohere rerank: PGVector
logged failure paths: 2

n8n
PGVector
OpenAI embeddings
Cohere
PostgreSQL

Opened from exported workflow files in my own instance - credentials do not travel with an export. The warning icons are missing keys, not broken nodes.

AutomationPrivate

An n8n canvas: build a Renderly payload, start video generation, extract the job id, wait 120 seconds, check job status, evaluate it, and loop back if it should continue polling. On completion an is-success branch splits into formatting the success and calling the auto-post workflow, or formatting a failure.

Is Success? - false goes to Format Failure

Async video publishing

Rendering takes minutes, so this polls for the result and branches on failure instead of hanging and hoping.

poll interval: 120s
networks published to: 6
outcomes, both handled: 2

n8n
Claude Sonnet 4.5
Renderly
Supabase
upload-post

Two node subtitles are blurred because they carried a server address. Opened from exported workflow files in my own instance - credentials do not travel with an export. The warning icons are missing keys, not broken nodes.

AgentPublic on PyPI

~/project $ axion -p "Find and fix the bug in auth.py"

$ pip install axion-code
$ axion login one time, saved to ~/.axion/credentials/
$ axion -m opus pick the model
$ axion --budget 1.00 cap the spend
$ axion --resume latest

Axion-Code

An autonomous coding agent that runs in the terminal, reads a codebase, writes code and runs commands.

model providers: 5
licence: MIT
published: PyPI

Python 3.11+
Anthropic
OpenAI
xAI
Ollama

Code PyPI

APIPublic

POST/api/video/generate/

202

{
  "job_id": "8f2c1a94-…",
  "status": "pending",
  "message": "Video generation started",
  "estimated_time_seconds": 240
}

Renderly

A Django REST API that renders video: Vertex AI Veo 3 for the B-roll, a HeyGen avatar composited over it.

job states: 5
accepted, then poll: 202
schema published: OpenAPI

Django REST
Vertex AI Veo 3
HeyGen
Celery
drf-spectacular

Code

ProductPublic

PayLink + PaylinkApi

The wallet screen. Balances and transactions are demo data.

PayLink

A fintech wallet: airtime, data, electricity, cable TV and exam pins, paid from a funded balance behind a transaction PIN and KYC.

API endpoints: 15
repos, one product: 2
auth with blacklisting: JWT

Next.js
TypeScript
Django REST
VTPass
JWT

Frontend API

APIPublic

GET/api/opportunities/

200

[{
  "id": "3b91f0c2-…",
  "title": "Software Engineering Intern",
  "type": "internship",
  "status": "open",
  "requirements": ["Python", "Git"],
  "required_documents": ["CV", "Transcript"],
  "application_deadline": "…",
  "applications_count": 0
}]

AspireBridge

A mentorship and internship platform for YABATECH - a Next.js frontend against a Django REST API, with real users.

API modules: 5
repos, one product: 2
schema published: OpenAPI

Next.js
Django REST
PostgreSQL
JWT
drf-spectacular

Frontend API

APIPublic

GET/api/chemical/

200

[{
  "id": "7d40ab19-…",
  "name": "Acetone",
  "quantity": 2.5,
  "molecular_formula": "C3H6O",
  "reactivity_group": "…",
  "chemical_state": "…",
  "hazard_information": "…",
  "location": "Store A",
  "expires": "…"
}]

Chemoventry

Chemical inventory management, where the domain model is the work: hazard data, reactivity groups, physical state and expiry.

fields per chemical: 9
repos, one product: 2
log on every change: Audit

Next.js
Django REST
PostgreSQL
JWT

Frontend API

AgentPublic

~/code-review-agent $ python src/main.py

$ git clone https://github.com/cypher125/code-review-agent.git
$ pip install -r requirements.txt
$ cp sample.env .env
$ python src/main.py serves the UI and the MCP server

Code review agent

Reviews a repository for quality, security and performance - and exposes itself as an MCP server so another model can call it.

languages analysed: 15+
callable as a tool: MCP
layer architecture: 4

Python
Gradio
MCP
GitHub API

Code

Skills

each links to its proof

Automation systems

17 systems, 831 nodes. The library, not a tool list.

Voice and conversational AI

The least contested lane I can actually deliver, and it is live.

Backends

The half most automation people cannot build.

Products

Applications people log into, not scripts.

RAG and knowledge systems

Chunking, embeddings and reranking - plus what happens when a document will not ingest.

AI engineering

Model picked per task, not one model for everything.

Contact

Tell me what you are trying to automate, or what you need built. Include the tools you already run on if you know them - it saves a round trip.

is the fastest way to reach me.