DirectorAlex

ModeHuman decides · agents code

Time zonesUS / UK / EU / Gulf

Open to a few conversations a year

00/Briefing

Alex Sokolov

Alex

Product leader, engineer by training

DirectorAlex designs, decides, directs
CrewAI agents write the code

AI is taking over doing things right. Doing the right things is still a leader’s job.

I find where AI agents pay for themselves, and help you build the first one.

I’m Alex, a product leader and engineer by training. AI agents are my engineering team: I design, decide and direct. I help CEOs, COOs, CIOs and CTOs pick the agent work worth paying for, name who owns each number, and plan the first pilot.

Alex Sokolov
Director Alex designs, decides, directs
Crew AI agents write the code

01/Telemetry Facts on record

CH-01Track record

18+years in IT

In management
14 yrs
Teams of
80+

CH-02Physics game

8 daysblank project to 9 levels

Test scenarios gating every step
103
Decision records kept
46
Creature designs run to calibrate difficulty
15,309

CH-03Persona engine

3memory layers in an open-source AI persona engine

Reasoning
inner monologue
License
MIT license
Tested by
AI judge

CH-04Simulation

17systems in a deterministic social simulation

Benchmarked social models
7 (AG1-a..g)
Automated Vitest suite
472 tests

02/Mission profiles 4 profiles

How I help

Not sure which format fits? Email one line about your company. I’ll suggest where to start, or tell you that you don’t need AI there yet.

MP-01Entry profile Start here

AI Opportunity Teardown

For
CEOs, COOs, CIOs and CTOs of SMB and mid-market companies who keep hearing “we need AI agents” and want a straight answer.
Problem
Vendors pitch tools. Nobody has mapped where agents pay for themselves, or whether you need them at all.
Book a Teardown by email

Payload

  1. 01
    A 90-minute working session with you and one or two of your leads, on the numbers you choose to share (NDA on request)
  2. 02
    A written memo within 3 business days:
    • your top agent opportunities, ranked by P&L impact and effort
    • a “Do we need AI here at all?” verdict for each one
    • an outline of the first pilot
Timeline1 week
FeeFixed

Quoted after a 30-minute call. 100% credited toward a Strategy Sprint if you book one within 30 days.

MP-022–3 weeks

AI Agent Strategy Sprint

For Leadership teams ready to commit budget and owners to AI.

AI programs get announced without a number, an owner or a date. Pilots stay demos and never reach daily use.

  • Agent opportunity map
  • KPI tree linked to the P&L, owners set by RACI
  • Model-agnostic pilot architecture
  • Vendor and build-vs-buy review
  • Rollout plan

Usually scoped after a Teardown

Discuss a Sprint
MP-032 weeks

AI Engineering Team Setup

For Founders and product teams who pay for AI tools but don’t yet ship product with them.

You pay for AI subscriptions but have no operating system around them. The agents produce output, not product.

  • Design-interview template, 200+ questions
  • Decision-record log
  • Test-gate checklist
  • Agent roles and handoffs
  • Cost guardrails

Set up on your first real project

Ask about setup
MP-041 day a week

Fractional CPO for AI Products

For Seed to Series A AI startups, and corporate AI units without a product leader.

You have strong engineers but nobody owns the product’s economics, so demos drive the roadmap.

  • Strategy and roadmap tied to a KPI tree
  • Pricing and discovery hypotheses tested on data
  • Team design and hiring criteria
  • Decision log
  • Weekly operating cadence

3-month minimum · live overlap with US Eastern mornings

Talk about a fractional role

I take on one or two client engagements at a time.Open to a few advisory, co-founder and C-level conversations a year.

All formats in detail

03/Method Operating procedure

Data first. Expert second. IT third.

Schematic 03-AFrom question to build

Gate 00Before anything gets built, five questions

  1. Q1What problem are we solving?
  2. Q2Does it exist?
  3. Q3Is it the one we thought?
  4. Q4Does it need solving?
  5. Q5What is the cheapest fix that works?
Stage 01FirstData
Stage 02SecondExpert
Stage 03ThirdIT
Off-ramp · before ITSometimes the answer is a process change and a spreadsheet.
P-01

Prove the problem before you price the solution.

Five questions come before anything gets built. Sometimes the answer is a process change and a spreadsheet.

P-02

Every initiative gets a number, an owner and a date.

A KPI tree linked to the P&L, with owners set through a RACI matrix. Without it, an AI program is a press release.

P-03

Decisions are the new bottleneck.

With agents, going from zero to a working MVP takes weeks. The value arrives only when leaders change roles and processes around the new tool.

P-04

Agents amplify management, good or bad.

If a manager can’t set a task and accept a result from people, agents won’t fix that. A team that plays together beats a team of stars.

RENTED PIERS? OLD CROSSINGNEW BRIDGEFORD UNTIL SOMEONE CLOSES ITBUILT OVERNIGHTNO BRIDGE NEEDED

The bridge rule

With AI you can now build a bridge overnight. But if nobody closes the old crossing, traffic stays where it was. If the river can be forded, you don’t need a bridge at all.

And don’t build on rented piers: today’s model prices are subsidized by investors and can change, so design for switching providers without a rebuild.

How I work with AI agents

04/Mission log 6 entries

Selected work

LOG-01B2B caseIndustrial pilot (under NDA)

AI Consilium

7 AI agents on the pre-investment committee: catching 'Watermelon KPIs' in 22 seconds

In large industrial groups, hundreds of shop-floor efficiency proposals and CAPEX requests wait months for committee review. Worse, initiatives that look green on a buyer's spreadsheet often hide massive TCO losses in maintenance, energy, or winter logistics.

A 7-role, 4-act multi-agent pre-investment committee: deduplicates hundreds of raw proposals, runs isolated parallel debates among 4 domain directors (FREE-MAD protocol with zero conformism), exposes hidden cross-department costs, and outputs a parametric TCO formula without exporting secret ERP prices to the cloud.

Role
My product concept, 7-agent cognitive architecture, mathematical deduplication, and deterministic scoring engine
Industry
Integrated heavy-industry group (name under NDA); directly applicable to manufacturing, mining, logistics, and retail
Protocol
FREE-MAD (Consensus-Free Multi-Agent Debate) + strict Physics-Cognition Boundary (scoring.py)
Security
Zero-Trust Split-Context — secret ERP prices and volumes never leave the enterprise perimeter

Why a CEO and CFO care

Eliminates "Watermelon KPIs"
one department's paper savings cannot pass at the expense of another's downtime
Separation of powers (LLM ≠ calculator)
agents uncover technical and operational risks, while deterministic Python computes all scores and FCF
Live without a 9-month ERP integration
the cloud committee outputs a parametric formula; plant economists plug in secret ERP volumes locally in 5 seconds
Auditable decision trail
every rejected or approved proposal keeps a full transcript of 4 domain critiques and the financial audit
800+ → 639raw shop-floor ideas compressed into unique hypotheses via two-pass DBSCAN
20–26 secfull 7-agent audit per hypothesis vs 400 hours of human committee meetings
~2,000×lower evaluation cost per 1,000 ideas using prefix caching and off-peak routing
0%conformism during debate via isolated parallel execution of the 4 domain directors

Why it matters A collective of specialized AI agents with strict separation of powers and zero conformism replaces hundreds of hours of pre-committee triage and blocks value-destroying projects before capital is committed.

Validated on a real pipeline of 800+ operational and CAPEX proposals at a major industrial group (name under NDA).

LOG-02B2B caseDemo on request

LLM Analog Finder

Find the substitute before you buy the duplicate

In large inventories the same kind of part often sits under different names, units and spec formats. Buyers can’t see the equivalent already on the shelf, so they order new stock.

An AI pipeline that cleans messy inventory master data, merges duplicates and finds functional substitutes for the item you need. Parameters first; a category-expert LLM steps in only where the algorithm is unsure.

Built
My idea and architecture. I wrote the first prototype, for cables, from scratch in Python and JavaScript
Scaled
In production at an industrial company. A team of engineers and AI agents scaled it under my lead
Tested
Open retail catalogs; synthetic inventory data in SAP and other ERP formats
76kcatalog items run through the pipeline at an industrial company
1 in 3items got a substitute or a conditional match
72%of “analogs” in electrical were duplicates of the same part
2 daysto onboard a new product group; the first build took 3 weeks

Why it matters One architect with AI agents builds a working prototype; a team scales it on a proven design. A new product group becomes configuration, not development.

Production numbers from an industrial company. Client name under NDA.

LOG-03Cognitive architecture · Open-source 1.0MIT license · Live on Telegram

Aristarkh Gradov

An autonomous AI showrunner & cognitive engine: 137k-char persona bible, self-curated news, and a live TV grid

Aristarkh Gradov is an autonomous 50-year-old media producer living on Telegram. In his broadcast channel, he runs a weekly TV schedule, selects his own news with exponential topic fatigue, and rewrites weak pitches on 'The Cutting Table.' In bot DMs, he remembers his morning posts at zero background token cost and runs a commercial diagnostic funnel.

A live proving ground for a 15-service cognitive engine: 153 pre-planning questions, 3 memory layers, an autonomous news curator (t½ = 3 days), a hidden Provocateur subagent, a 3-layer Clean Room IP filter, and a 'Velvet Rope → Cold Shower' monetization funnel.

Persona depth
153 design interview questions (resolving 16 contradictions), a 137,000-character persona bible, and 19 voice-calibrated behavioral situations
Autonomous curation
curator.py scores 10 news items (0..10) with exponential topic fatigue (3-day half-life) and skips the slot if nothing scores ≥ 6
Anti-cliché & IP shield
Sliding-window motif quotas (variety.py) + 3-layer Clean Room IP biography filter in ChromaDB at 0 LLM tokens
Broadcast bridge
Unified channel-to-bot memory at 0 background tokens (recent_posts.json) + Velvet Rope deeplink funnel

How the same engine works in your company

Memory architecture (3 layers)
corporate policies, client interaction history, and extracted project facts
Two-pass reasoning + Provocateur
private inner monologue plans the angle before a single word is sent to the user
LLM-as-a-Judge & surgical micro-edits
automated quality, tone, and IP guardrails that fix single sentences without ruining cadence
Cost router & 0-token state sync
predictable unit economics per interaction without runaway background summarization bills
137,000characters in the persona bible, speech patterns, and hard constraints
0 tokensbackground cost to synchronize channel broadcasts with 1-on-1 bot DMs
t½ = 3 daysexponential topic-fatigue half-life in the autonomous news curator
18 rolesin the 3rd-person diagnostic mirror (replacing preachy 2nd-person coaching)

MIT-licensed 1.0 core · reply guard enforces honest AI identity · live broadcast grid + 1-on-1 diagnostic bot

LOG-04Social simulation & Decision SandboxDeterministic engine · 472 tests

The Rat Box

A computational social time machine: 17 systems, a 15 ms parallel-universe forecaster, and a 3D documentary studio

In organizations and policy, major reforms are usually tested on live people by trial and error. The Rat Box is a deterministic multi-agent decision sandbox set aboard a generation ship, where centuries of policy trade-offs, faction splits, and cascading failures play out in seconds.

A 17-system tick pipeline, 7 social-science models with automated benchmark tests (AG1-a..g), 1,000-year bit-for-bit determinism across 4 JS engines, a 15 ms parallel-universe 30-day policy forecaster (forecastLever), a 96.8% Causal Inspector, and a built-in 3D documentary filmmaking studio (/?studio).

Scientific core
17 simulation systems and 7 social-science models with automated benchmark tests (AG1-a..g), including network SEIR within ±6% of analytic curves
Decision sandbox
46 indirect levers across 5 domains + <15 ms parallel-universe 30-day forecaster (forecastLever) + Causal Inspector (96.8% root-cause coverage)
Determinism
Bit-for-bit identical state hash after 1,000 years (365,000 ticks) across Node, Chromium, Firefox, and WebKit; 0.30 ms/tick at 10,000 agents; 0 LLMs in runtime
Ship v2 & Studio
'Drum on a Spine' engineered from 179 NASA BVAD / Kalpana One sources, Style G entropy decay, 11-part TABS figures at 1.9 ms/frame, and the /?studio director toolkit

Why executives and systems architects care

Pre-decision counterfactual forecasting (forecastLever)
see the 30-day net delta of any policy in a cloned parallel reality before committing
Exposing cascading policy failures
why a "30% ration saving" or a shift-priority bug spirals into legitimacy collapse and shutdown
Zero black-box outcomes
every model passes an analytic benchmark test, and the Causal Inspector traces every crisis to 5 root causes
Two-Door architecture (Game + 3D Documentary Studio)
the exact same simulation engine powers the documentary series "Chronicles of the Ark"
17 / 472tick-pipeline systems / automated Vitest tests gating every commit
< 15 msto clone the world twice and forecast 30-day policy consequences
0.30 msruntime per simulated day for 10,000 agents (deterministic over 1,000 yrs)
96.8%of crises decomposed into 5 exact causal blocks by the Causal Inspector

Why it matters Before you trust a business agent or policy model with real capital, it should clear the same bar as The Rat Box: an automated benchmark test for every model, 100% replayable runs, and a transparent causal tree.

Discuss applying this simulation & sandbox approach Content note: deals with mature themes (civil conflict, institutional collapse, mortality).
LOG-053D naval engineering & roguelikePrinceps Games · Steam — Dec 2026

Hullcraft

Graphics are 20% of AI gamedev. The other 80% is the 'WHAT IF?' dramaturgical & physics engine.

Today, anyone with Claude Code or OpenAI Codex can vibe-code a basic 2D/3D game in Godot, Unity, or Unreal in 24 hours, or ship an MVP in 2–5 days. Hullcraft is about what happens after day five: building the 80% systemic and narrative engine that turns an AI tech demo into a real game.

A 3D naval engineering sandbox and roguelike voyage across an 1899 Admiralty chart of the Shetland island of Hamarsey — built from scratch by a 1 + 1 pair (1 Product Architect + AI Agents) across 66 systems and 704 automated tests.

Core thesis
Visuals and shaders are 20% of the work; 80% is the dramaturgical and systemic gameplay engine (story + WHAT IF? variability)
Architecture
66 decoupled gameplay systems, 704 automated tests, 20 custom GLSL shaders
Physics & Hull
Dual-Grid Wave Function Collapse hulls (256 configurations) + 65 physical parts (impulse-theory propellers, cavitation, gearboxes, sails)
Audio & World
27 real-time synthesized sound spectra (0 external .wav samples), procedural village Kirkvoe, and 6 GPU sea monsters up to the 120m Stoorworm

Why game producers and CTOs care

Beyond the 48-hour vibe-code trap
how to architect deep systemic replayability and narrative arcs with AI agents
Self-verifying procedural generation
code-enforced collision proofs (78 cottages, 139 stone walls, 0 overlaps)
Headless AI playtesting (--test-voyage)
agents sail the route in headless mode and tune physics from JSON telemetry
1 + 1 production velocity
one product architect + AI agents delivering studio-grade systemic depth
20% / 80%visual surface vs dramaturgical & systemic gameplay engine
66 / 704decoupled systems / automated tests gating every build
65 / 20physical ship parts / custom GLSL shaders (0 external audio files)
120 mprocedural Stoorworm sea dragon awaiting in sector H12

Why it matters A full 3D systemic game with real hydrodynamics, WFC geometry, and 704 automated tests is harder than 90% of enterprise software — proving what a disciplined 1 + 1 workflow can ship.

Steam page launching December 2026 · press & demo build available on request

LOG-06Physics gamePrinceps Games

Evolution Workbench

A physics game I designed and directed. AI agents wrote the code.

The agents also helped with the in-game text. Build an alien creature from parts and send it across hostile worlds. When it falls, an “autopsy protocol” tells you why, in numbers.

9 levels / 8 daysblank project, a from-scratch rewrite after a first prototype
15,309creature designs run through the physics to calibrate difficulty
103automated test scenarios gate every step
46decision records · no third-party assets

Why it matters The decision log and test gates behind this game are the system the AI Engineering Team Setup puts in place.

Browser demo coming to itch.io · desktop browser only

All projects: business cases, agent lab, games

05/Transmissions Manifesto

  1. TX-01

    Managers do things right; leaders do the right things.The line goes back to Warren Bennis. AI is about to take over the first half.

  2. TX-02

    No P&L number with a name and a deadline attached means no transformation. Just a press release.

  3. TX-03

    Code is no longer the bottleneck. Decisions are.

  4. TX-04

    If you can’t manage people, you won’t manage agents.

  5. TX-05

    Sometimes the right answer to “which AI do we need?” is “none, use Excel.”

06/Briefings 6 talks

Speaking

Keynote · panel · fireside chat · workshop · in English (C2) · remote worldwide · in person by arrangement

Speaker kit: bio in 3 lengths · photo

  1. T-01Robot Managers Do Things Right. What AI Agents Won’t Replace.
  2. T-02My Engineering Team Is AI Agents: A CPO’s Operating Model.
  3. T-03No P&L Number, No Transformation.
  4. T-04Building an AI Persona That Thinks Before It Speaks.
  5. T-05The One-Person Company: Running a Business with AI Staff.
  6. T-06Sometimes the Right AI Answer Is “Not Here.”