Carlo Perito

AI-first software engineer

open to conversations

I build AI products that cite their sources and ship with agent fleets.

COMING SOON · AN AI VERSION OF CARLO

Ask about Carlo's work.

Every answer cites its source.

Hi, I'm an AI version of Carlo, built by him. I answer only from material Carlo has approved and link every claim to its source. I can't speak for him on salary, start dates or availability, and if something isn't in his material I'll say so.

Try asking

How does Carlo's agent loop catch bugs?What has Carlo shipped end to end?How does cited retrieval work in MemReez?

Your questions are held with a random session ID, never your IP address or browser details. If you agree, Carlo keeps the conversation for up to 180 days to improve AI Carlo; otherwise its text is deleted within 3 days. Privacy details

How this works

Carlo approves every item that goes into AI Carlo's material.
Answers are written only from that material, and every claim lists its source.
If the material doesn't support an answer, AI Carlo says so and invites you to contact Carlo.

ENGINEERING OS

How I build now

Measured on MemReez · 2026-09-18–2026-09-24 vs 2026-09-25–2026-09-30

The agent loop

One agent writing and checking its own work is grading its own homework.

An orchestrator plans and reviews, executors build, and five separate layers check each change before it merges.

freeze audit · Codex · spec gaps before code
warm review · Orchestrator · side doors, cold starts, visuals
probe runner · Sonnet subagent · race and interleaving probes
cold review · Codex · the exact await that goes stale
CI · GitHub Actions · skipped local steps

Median hours first commit → merge: 10.3 (Sep 18–Sep 24) → 3.75 (Sep 25–Sep 30)

Experiments

Choosing models and effort by feel burns quota.

Head-to-head runs decide a tier change before it is adopted.

cheaper model as red-tier reviewer → green tier only
written probe templates vs the cold reviewer → templates raise recall; cold review stays the gate
fast model vs Sonnet as probe runner (2 rounds) → Sonnet stays the probe runner
executor effort xhigh -> high → high for red tier, cheaper model for green
reviewer top tier -> mid tier → mid tier high effort for audits and cold reviews

every experiment is one run per arm; no noise baseline yet

Engineering under limits

Weekly quotas and CI minutes are hard ceilings.

Each limit got a standing response instead of a workaround.

reviewer weekly quota → mid-tier reviewer, top tier only for disputes
executor monthly quota → no xhigh; cheaper model for green work
CI minutes → draft skip + local gates + one billed run

Web gates failure rate: 31.6% → 8.8%

Tokens per task PR (executor): 3.4M → 1.9M

Tokens per task PR (reviewer): 820K → 420K

Slide 1 of 6: The agent loop

Projects

MemReez
web · iOS · Android
Hey Memories — cited answers over your own memories
VaultDoc — document provenance on a blockchaincoming soon
Law OS — state-by-state legal retrievalin development

Featured

My Story

From aerospace to reliable AI autonomy. A scroll-driven journey through 20 years of building machines and code.

Explore my story →

Services

Web front endMemReez web — Hey Memories chat, citation chips, and the warm light UI visitors use today.
Back end & APIsMemReez ingestion and retrieval APIs behind grounded answers over approved material.
iOS & AndroidMemReez ships on iOS and Android alongside the web app, same cited-retrieval story.
Deploy & hosting

Experience

Helicopter mechanic
Engineering tech
Full-stack (Mimo cohort, 2022)
AI-first solo founder

Contact

Prefer a human?

Contact Carlo cp4Programs