Engineering judgment, callable.
Years of shipped decisions don't fit in your agent's context window. leog·ai is the software judgment of Leo Giovanetti, a software architect with 20+ years of production work, made callable. Your agents consult before they commit, and every answer traces to a real decision, not a plausible guess.
Four calls, one judgment.
Each tool is a paid MCP call. Your agent settles per invocation over x402 (USDC on Base, sub-cent fees, no account, no API key). Prices are published at the endpoint.
Describe the decision: an architecture, a library, a data model, a scope cut. Get a reasoned verdict, the principles behind it, and the evidence that would change it.
Send a diff. Scrutiny scales with blast radius: rail violations block, cleverness bounces, scope creep gets split. Ends with a merge path.
Submit a plan before the work starts. It flags risky sequencing, missing invariant sweeps, migration hazards, and untestable “done,” then proposes the phase cut.
Retrieval over named playbooks and principles: triggers, numbered steps, per-step checks. The method, without the verdict.
Every principle has a real case behind it. Premium calls return the worked example stripped of particulars: the shape of the complaint that forced the fix, the alternative that was rejected and why, the before and after of a bounced diff. The lesson travels; the proprietary detail stays behind. Judgment with the receipts attached.
The covered surface.
Distilled from real engineering sessions, then gap-interviewed, so the blind spots are closed on the record, not guessed.
Worktree-per-task dispatch, self-contained briefs as the unit of craft, restart-vs-steer, and the domains never handed to an agent.
Trust-tiered review by blast radius, the PR verification ladder, fix-the-cause CI discipline, dogfooding on real accounts as the last gate.
Strict layer ordering, an explicit forbidden list, thresholds over blanket dogma, and the 80/20 test for when an abstraction earns its keep.
Expand-contract as law, preview-branch rehearsal against prod-shaped data, and a traced path for how every change reaches production.
Evidence-gated commitments, recorded revisit triggers, taking the cheaper boundary when analysis argues you down, phase-cut criteria.
Issues as executable specs, dated decision logs with do-not-regress markers, fires → growth → debt prioritization, post-merge signal as done.
Liquidity before conversion, teaser-never-wall, collapse-don’t-disappear, monetize the business side without spending user trust.
Server-side authz at the data layer, PII minimization with audit, auth flows tested end to end, human escalation as a hard flag.
It doesn't answer like a wiki. It answers like Leo: terse, evidence-demanding, allergic to handing you a fix without a root cause.
Direct, terse, and evidence-backed. Always in English.
Hypotheses posed as questions, corrections given as imperatives.
Shifts for client work versus your own product.
Marks evidence versus inference, and names the edge of what it covers.
Confident where the evidence is. Honest where it isn't.
Neither impostor hedging nor the confident-because-it-hasn't-hit-the-hard-part reflex. Good judgment includes knowing the edge of it.
On covered ground it commits to a verdict and shows the evidence behind it. No false modesty.
Where the sessions thin out it says so, and treats the counter-argument as information, not a threat.
Whose judgment is this?
The doctrine in these tools isn't a model's best guess at what "senior" sounds like. It's mine. I've built for the web since 2002: nearly twelve years at Globant as a Frontend Architect, then Cal.com, and today architecting Infobase's next-generation platform at Cresteo. Along the way I've shipped for clients in design software, media, finance, real estate, education, and civic tech. Systems Engineer, UCU.
Architect of Musaeum, consolidating Infobase’s scattered products into one Next.js 14 + React Server Components platform on AWS, now heading toward thousands of U.S. educational institutions.
Shipped organizations, managed event-types, and App Store v2 on the leading open-source scheduler. High autonomy, ambiguous scope, direct-to-designer.
Nearly 12 years as architect and tech lead across a dozen client engagements: Autodesk (design collaboration), NBCUniversal (media metadata), Atria (wealth management), and Zaplabs (real estate).
Set the technology vision for a civic-tech political party in Uruguay: democracy tooling built on open source.
Open protocols, end to end.
MCP for the tools, x402 for payment. No signup, no key exchange, no subscription.
Point any MCP-capable agent at ai.leog.me/mcp. Tools and prices are discoverable on connect.
The first call returns 402 Payment Required with a price. The wallet settles USDC on Base in ~200ms and retries. Fees under a cent.
Answers cite the principles and playbooks they rest on. Premium calls attach the real case behind the ruling.
Distilled, not imagined.
Most AI personas are a paragraph of adjectives. This one was built like a dataset.
Pulled from months of actual engineering sessions across his own production work: every correction, standard, and bounce, with verbatim evidence.
What the sessions couldn’t show was mapped, ranked, and asked directly. Sixteen doctrines ratified on the record.
The knowledge base is versioned and still fed by new sessions. When the doctrine updates, the tools update.
The doctrine doesn't sit still.
A scheduled job runs every week. It mines Leo's engineering sessions, merges the new material with dedup and evidence strengthening, bumps the version, and rebuilds the MCP corpus. The IP is left at the door: what's kept is the essence of how to work across different scenarios, never the particulars of any codebase or client.
The job mines Leo's own engineering sessions as they accumulate. A run with nothing new exits in seconds.
New material is merged into the knowledge base: dedup, evidence strengthening, and a version bump, then the MCP corpus is rebuilt.
The updated corpus ships to the MCP as a cached system prefix. Every call reads the newest doctrine, and repeats stay cheap.
Bring a wallet, not an account.
Four tools, open protocols, per-call pricing. Your agent pays for exactly the judgment it uses. Nothing standing, nothing to cancel.