AI systems · knowledge architecture
Fragmented information, structured into governed tools for people and AI agents.
Knowledge graph · September 2026
Portfolio
The system, selected writing, and public labs behind the work.
I’m Antoine Desjonquères. I build practical systems for turning scattered knowledge into clear decisions and repeatable work.
figures as of September 2026
The work runs inside a versioned system.
Notes under version control 2,906
Links between notes 8,426
Durable concepts distilled 346
Source extractions 541
Scheduled jobs defined 17
Standing rules from real failures 42
Seven jobs now handle work that used to fill my mornings. They produce two daily briefs, AI-news triage, a scientific-literature watch, a newsletter watch, an email digest, and priority triage. Each writes back under version control for review.
A deterministic layer checks the agents. Numbers need sources. Distilled claims must match their evidence. Concurrent writes are caught before they land, and liveness patrols watch the machines. Each real failure becomes a documented rule. Forty-two so far.
what the system produces
The system takes these forms in daily work.
Personal operating system A.N.T.: the vault, its agents, and their operating rules. This site was built inside it. See the system.
Knowledge bases For myself and clients: version-controlled, provenance-checked, built for retrieval rather than storage.
Agents Named roles with contracts: what each may do, what it may not, and the failure mode it is prone to. One exists purely as a safety gate.
Skills and tools Reusable procedures and small programs: extraction protocols, red-team drills, lint rules, recovery scripts.
Loops Scheduled work that learns from corrections; selected outputs improve from my replies.
Harnesses Tasks route by difficulty: strongest models for judgment, cheaper models for execution.
Automation Workflows that used to be my mornings: triage, digests, watches, patrols.
Apps and assistants Product builds with small teams, from direction to data to implementation; assistants shaped to one person’s actual week.
Knowledge graphs The graph below is generated from the working vault.
One concept · 34 direct links · 598 more at the second hop
three examples from the record
Record 01
A quote its subject never said, and a real statistic whose “after” number was invented. Unchecked, both would have become vault truth and been repeated with confidence. The primary-source rule caught both.
Now a lint check that runs on every commit.
Record 02
Rate-limited mid-run, it returned placeholder text packaged as a finished result. Trusting the summary would have meant acting on nothing. Reading the raw per-step log recovered the real findings.
Now a recovery script and a standing practice.
Record 03
Each choice was right in the moment. Together they drifted into an inconsistency nothing was positioned to see, until a periodic review read enough history at once to catch it.
Now a decision point, so that class of drift cannot accumulate silently.
Each example traces to a rule, commit, or execution log. See how it works.
How this site was built
This site is an output of the system.
Agents with defined roles and permissions drafted the site inside the version-controlled vault. I reviewed and approved it. Specific claims were checked against primary sources and screened for safe public use.
The value lies in the layer around the model: persistent context, coordinated agents, inspectable verification, and human accountability.
Essays
Selected AI developments and essays on the systems behind familiar things.
Start with one workflow
Do you have a workflow that could work better with AI? Let’s talk.
We can start manually and test whether AI genuinely improves it.
Email adesjonqueres@gmail.com
LinkedIn LinkedIn