Available for freelance & contract work
Dany Chausse
AI Automation & Technical Writing
I build AI automation — cold-outreach engines, video pipelines, local agent stacks — and write the documentation that survives contact with the machine.
Every piece I ship comes with the code it describes, the log line it actually wrote, and a test that fails if the claim stops being true.
Proof — systems built, then written up
Built & verified
-
AI automation · Python · SMTP & IMAP
A cold-outreach engine that proves each email actually left
The sender behind the agency’s own outbound. A pre-flight duplicate gate checks the SQLite log before anything is queued, every SMTP stage is recorded, and the stored Message-ID is read back out of the database after each send — so “sent” is a record, not a print statement. Attachments are hashed before the MIME part is built, the warm-up cap is enforced in code, and a matching IMAP watcher files each reply against the lead that earned it.
Python · 683 lines · SQLite send log · dry-run against a local SMTP sink before a real inbox
-
AI automation · Python · yt-dlp & ffmpeg
One long video in, three captioned vertical clips out
A clipper built for the short-form side of the business: hand it a YouTube URL and it returns 9:16 clips with captions burned on the frame — yt-dlp fetches, faster-whisper transcribes on CPU, ffmpeg cuts, reframes and burns the cues. Proved on two public sources, a 13-minute and a 10-minute 1080p video, three clips each in 125.6 s and 117.9 s of wall-clock. Both known defects — a caption baseline that collides with baked-in lower thirds, and a centre crop that loses the edges of screen recordings — are written down next to the code rather than left for a client to discover.
Python · 627 lines · 1080×1920 at 30 fps · audio and burned-in captions verified on the exported files
-
AI agent stack · local models
A guard that sits between the model and the machine
Tool-calling is where a small local model does the most damage: it calls something that does not exist, or returns an empty result that reads exactly like “nothing found”. The guard validates every call against a JSON schema before it executes, keeps a failed tool distinguishable from an empty one, and ships with the terminal output each port actually produced — Node and Python, run against a local Qwen and DeepSeek bench.
Node.js & Python · JSON-schema validation · run transcripts kept with the code
Published technical writing
-
Tutorial · Node & Python
Build a Tool-Call Guard for Local Models in 58 Lines of Node
A zero-dependency wrapper that makes “the tool failed” impossible to mistake for “the tool found nothing” — plus a six-case harness that proves it on your own machine, and the same guard ported to Python.
-
Case study · reproduced failures
The Hidden Cost of Small Local Models in Agentic Workflows
A local 4B model has a marginal cost of zero — and a bill that moves to a line item most agent tooling never logs. Two failures, reproduced command by command on one small task, compared against a hosted model.
-
Essay · pricing analysis
The AI Discount Is Really a Testing Discount
A $400-versus-$700 rate card looks like a verdict on machine-written prose. Read closely, it’s a price for verification — which changes what a technical writer should be selling in the first place.
About — AI automation & technical writing
I build AI automation and write the technical documentation that keeps it usable. Most of my work is one of three things: an automation that takes over a manual routine, a lead-generation engine that finds and verifies its own prospects, or the documentation a new operator can follow without calling me.
Every engagement ends with a measurable result — hours returned per week, outreach volume a log can prove, or a run that either passes its check or fails loudly. I write down what the system actually does, including the parts that did not work the first time.
-
Writes about
AI automation, LLM tool-calling, local inference, developer documentation
-
Builds with
Node.js, Python, local model runtimes, CLI tooling, test harnesses
-
Working style
Reproduce first, then write. Every claim carries its evidence.
-
Available for
Automation builds, lead-generation engines, technical documentation
Contact — start a project
If you need an automation built, a tutorial programme, a docs rewrite, or a case study that holds up under a reader trying to run the code — I take on freelance and contract work, and I’m glad to start with a short paid trial piece.