Search
Start AI Tools with Intent Solutions
The implementation journal of Intent Solutions. Follow the work of turning AI capability into reliable production systems: the builds, evaluations, failures, and operating lessons your team can use for itself.
Curated long reads
Let the Model Judge. Make the Code Decide.
The model owns judgment. Deterministic code owns the gates, the commit, and the audit trail. How I rebuilt a daily writing pipeline so a hung session can no longer brick tomorrow.
Adversarial Review: The Six Lenses That Halted a Rollout
Six lenses, live state, eighteen risks. I thought the safety work was shipped. The review proved three assumptions were theater. The durable rule: the model proposes; the system owns identity, trust, and state.
The Moat Is the Trust Layer
A local-RAG demo became Intent NEXUS. The moat is not the chatbot. It is egress control, code-enforced citations, a ledger you can check, injection defense, and evals that cannot lie quietly to you.
Recent posts
When a Third-Party CI Credential Expires, the Deploy Fallback Has to Already Live on the VPS
A VPS-resident deploy script plus a SHA-pin guard turns any third-party CI credential expiry into a one-line SSH recovery.
ReadWhen git describe Traps Your CI Semver Bump, Look Up the Highest Existing Tag Instead
A bot's [skip ci] version-bump commit that was rejected by branch protection creates an annotated tag outside the branch's ancestry. git describe keeps returning the older reachable tag, the bump recomputes a version that was already tagged, and the release job fails with `tag already exists` on every push. Use `git tag --list 'v*' --sort=-v:refname | head -n1` (highest existing tag, not git describe) and the version stays monotonic forever.
ReadWhen Your Documentation Evaluation Is Driving the Cleanup, Report Gap Answers by Rank Instead of Adding Them to the Scored Set
A self-improving retrieval evaluation should not grade its own homework. Report the rank of each gap-filler doc in retrieval (answerRank per gap) and leave the fixed scored question set untouched.
ReadCoordinating a Product Launch Across Multiple Repos
The order of operations for a product launch across three repos and three releases: what each step prevents and what to check after the merge.
ReadVerify Audit Evidence with a Signed Payload, Not by Format
When eight audit-evidence publishers each emit a different format, one canonical signed payload gives one verifier a tampering signal without erasing history.
ReadSMTP 250 is not delivery: three receipts that fix transactional mail
SMTP 250 is not delivery. Three receipts fix transactional mail: a frozen ApprovedMessage, phase-classified outcomes (RFC 5321 §4.2.1), and reconciliation before any resend.
ReadFirst Slice in One Day: Eight PRs, One Sign-Off
Split a first slice into one PR per layer, capped by a governance PR with the sign-off. Each layer, the merge-guard hiccup, and the day the slice landed.
ReadSeries & curriculum
MCP for Beginners
End-to-end Model Context Protocol curriculum — solutions in Python, TypeScript, Java, Rust, C#, .NET — translated across six languages.
Agentic Design Patterns
Patterns and anti-patterns for production agent systems. Decision frameworks, prompt scaffolds, evaluation harnesses.
Tiny Recursive Models
Building small recursive systems where simple loops compose into emergent behavior. Series in progress.
Research & Curriculum
Long-form research articles, learning paths, and reading lists for the AI builder community.
Retrospectives
August 2026: A Control That Fires on Every Case and Changes No Outcome Is a Log Line
A control that fires on every case and changes no outcome is a log line, not a control. August 2026 retrospective: 31 of 31 posts flagged for inflation with the distribution unmoved, a learned cap rule that caught zero, and a rubric six weeks stale that still names a retired rule.
July 2026July 2026: Every Instrument Reported Success, and One of the Published Posts Was a 404
July 2026 retrospective: Tier 1 collapsed from 46% to 17%, the anti-inflation rule adopted July 3 fired zero times all month, the grader demoted half the catalog on a title keyword list, and a post with a full green audit trail was never published at all.
Let's build something that ships to production
Intent Solutions is an applied AI engineering company. We build with your team, test against your actual work, and transfer the knowledge needed to operate and evolve the system. This journal shares the methods and lessons that can be made public.
No commitment required. We respond within 24 hours.