AI production systems
Acceptance tests, evals, MCP permissions, observability, cost ceilings, approval gates, trace replay, recovery, rollback, ownership handoff.
Operator console · LiveControl planes for AI agents.
Control planes for AI agents.
Not slideware.
Policy gates, immutable audit, eval harnesses, and budget ceilings. The operating layer that takes blocked Claude and agent initiatives to production.
Agents are downstream of workflow intelligence. Connect the source systems first: files, tools, reports, inbox, leads, services, logs. Then the agents have something real to read.
MCP servers
2 prod
Providers
OpenAI · Anthropic · Google
Base
Columbus, OH
Two lanes. Pick the proof you need.
Airtable exit and rescue
Dependency inventory, linked-record preservation, automation mapping, permission parity, reconciliation, cutover, rollback, Postgres and Supabase patterns.
The operating layer, not the demo layer
Wire the source systems first →
Files, tools, reports, inbox, leads, services, logs. All connected into a queryable map before a single agent is deployed.
Ship agents with guardrails →
Policy gates, per-caller budgets, and immutable audit are wired into every tool call. Not bolted on after review.
Prove it with telemetry →
Every claim comes with a log, a cost line, and a rollback path. Logs over vibes. Production over theater.
The proof standard on this domain
Every project page carries the same five fields. A field that cannot be filled from a public artifact is marked as not published rather than estimated.
1
Repository or artifact
A link to the source, the package, or the running product. Private builds are labelled private and claim less.
2
Architecture
The planes, the boundaries, and where control sits. Written from the code, not from a pitch deck.
3
Tests and gates
What runs in CI, what has to pass before a change lands, and what the suite does not cover.
4
Known limits
The parts that are prototype, unfinished, or deliberately out of scope, stated on the page itself.
5
Last verified date and classification
Open source, internal, experiment, sanitized example, fork or evaluation copy, or private build. Each with the date the facts were read from the repository.
Publicly checkable
Counts you can verify yourself.
Every number here resolves to a public artifact: a GitHub repository, a published MCP server, or a project page on this site. Dashboard-only figures are not published.
22
public repositories
classified in the ledger
10
project proof pages
repo, tests, limits
2
MCP servers
published and installable
13
years enterprise IT
US-based
Built in the open,
wired for production,
not slideware
Active builds behind the operating layer. Each project page carries the repository, architecture, tests, known limits, and a last verified date.
Open Source Systems
Web Products & Interactive Experiences
Reference patterns
Sanitized problem, intervention, verification.
These are reference patterns, not client quotes and not case studies. No client name, data, timeline, or outcome figure is published on this domain.
Agent work shipped without a definition of done
AI production systems
Reference pattern
Runs were reviewed by reading transcripts, so plausible output passed as correct output.
Intervention. Executable acceptance tests written before each run, split into task assertions and invariant assertions, with the verdict stored against the trace ID.
Verification. Re-run the suite on the pre-change commit and confirm it fails for the right reason.
Read the pattern →Tool calls with no permission boundary
AI production systems
Reference pattern
An MCP surface exposed every tool to every caller, so a read-only assistant could reach write paths.
Intervention. Typed capability surface per caller, per-caller budgets, and structured logs on every call.
Verification. Attempt a write from a read-only caller and confirm the denial is logged, not silently dropped.
Read the pattern →Airtable base carrying production dependencies nobody had mapped
Airtable exit and rescue
Reference pattern
Automations, scripts, and external integrations referenced tables that a migration would have silently broken.
Intervention. Full dependency inventory before any schema work, followed by automation and script mapping to explicit targets.
Verification. Every inventoried dependency has a named owner and a post-cutover check.
Read the pattern →Linked records flattened by a naive export
Airtable exit and rescue
Reference pattern
A CSV-based migration path dropped link direction and multi-link cardinality.
Intervention. Preserve links as join tables with stable external IDs, then reconcile counts on both sides.
Verification. Row and edge counts reconcile between source and target before cutover, not after.
Read the pattern →Architecture
The systems behind the engagements.
Production patterns, not theoretical designs.
TechTide Swarm 357 — Multi-agent orchestrator
Single runs utilizing up to 50 concurrent worker nodes. CEO agent delegates to ten sub-orchestrators, each managing five specialized workers.
┌──────────────────────┐
│ CEO Agent │ LangGraph
│ (delegation tier) │
└──────────┬───────────┘
│
┌───────────┬───────┼───────┬───────────┐
▼ ▼ ▼ ▼ ▼
┌───────┐ ┌───────┐ ... ┌───────┐ ┌───────┐
│ O1 │ │ O2 │ │ O9 │ │ O10 │ Claude Agent SDK
└───┬───┘ └───┬───┘ └───┬───┘ └───┬───┘
│ │ │ │
5 workers 5 workers 5 workers 5 workers = up to 50 concurrentStripe Connect V1 → V2 — Thin-slice migration
Dual-API webhook coexistence. V1 and V2 endpoints run in parallel; feature flags control traffic routing per merchant during rollout.
Stripe Webhooks
│
┌──────────┴──────────┐
▼ ▼
/api/stripe/v1/ /api/stripe/v2/
webhook webhook
(legacy merchants) (migrated merchants)
│ │
└─────────┬───────────┘
▼
Feature Flag Router
(merchant_id → version)
│
▼
Unified Business Logic
(version-agnostic)Discoverable endpoints
MCP servers shipped to production.
Model Context Protocol servers registered on major MCP directories and installable via the Anthropic CLI.
techtide-mcp
TechTideOhio/techtide-mcp
Production MCP server for TechTide AI workflows. Exposes Claude Code skill orchestration, Lovable project management, and n8n workflow triggers.
cipherclaw-mcp
Alexi5000/CipherClaw
Debug-first MCP server for OpenClaw agents. Causal debug graphs, cognitive fingerprinting, predictive failure detection, and soul integrity checks.
Operator stack
The tools behind the operating layer.
The actual surfaces I build, ship, and operate from. Not a menu — a working setup.
Build & code
Cursor
Primary IDE. Repo-wide context, agent mode.
Claude
Reasoning and long-context planning.
Codex
Diff-scoped edits and refactors.
Lovable
Full-stack prototype surface.
Operator agents
Zo Computer
Local agent runtime with connected app context.
Manus.ai
Autonomous task runs against real tools.
Granola
Meeting capture into structured notes for downstream agents.
Creative & docs
Canva
Static brand and social artifacts.
Gamma
Long-form decks and briefings.
CapCut
Video edits and demo captures.
Infrastructure
Railway
Long-running workers and services.
Supabase
Postgres, auth, and RLS. Primary data plane.
Cloudflare
Workers, edge routing, DNS.
GitHub activity
6,326 contributions across 120+ repositories.
Primary accounts: @Alexi5000 (personal) and @TechTideOhio (organization).
6,326
contributions / year
public graph
120+
repos touched
public graph
2
GitHub orgs
@Alexi5000 · @TechTideOhio
FAQs
It is the technical proof layer behind TechTide AI. Implementation patterns, project pages with repositories, tests and limits, and a classified ledger of public repositories. The person, story, and speaking context live on alexcinovoj.com. Commercial scope and pricing live on techtideai.io.
Explore the console
Read the proof. Then talk scope with TechTide AI.
This domain publishes implementation proof only. Commercial scope, deliverables, and pricing live on TechTide AI: AI Production Systems for agent and LLM work, Airtable Exit and Rescue for moving off Airtable.
alexcinovoj.dev
© 2026 Alex Cinovoj · TechTide AI · Columbus, OH