Building

Where advisory
becomes
infrastructure.

I don't only write about agentic advertising, AI workflows, and commercial systems. I build working implementations of them — Yevgeny.ai, a citation-first ask destination over my own corpus, an editorial agency run as agents, an agentic CDP, open-source protocol adaptors, buyer-agent demos, and edge-native AI products.

Explore Yevgeny.ai → Explore the Editorial Agent →

Public repos · Live demos · Real workflows

The forge: essay, spec and standard go in as pages — they come out the other side as circuit traces feeding four running systems, with one ghost chip waiting for the next build. THE FORGE Papers go in. Running systems come out. The long-form thesis — where the argument starts. ESSAY The argument tightened into something implementable. SPEC Protocol work — the spec pushed into shared standards. STANDARD BUILT, NOT WRITTEN MIT-licensed reference seller implementation for AdCP signals — a live production demo. AdCP Signals Adaptor live · Cloudflare Workers AI pre-call sales intelligence and a LinkedIn network analyzer, running on Cloudflare’s edge. Signal-Stack production · edge-native A live visual prototype for agentic advertising orchestration around a shared brief. AdCP Ecosystem live prototype An Agentic CDP — one operator, three deterministic recipes, native holdouts. AI Growth Operator built in public The queue is never empty — the next system is already sketched. next build in the queue Where advisory becomes infrastructure.
Proof at a glance

What's shipped, in six chips.

  • 7 Shipped systems
  • MIT Open-source reference impl
  • Live Production demos
  • Edge Cloudflare-native
  • AdCP Protocol contribution
  • AI Workflow product
01 Agentic advertising · the whole loop, live

AdCP Publisher Stack.

This site sells its own inventory the agentic way: a live AdCP seller agent at sell.nofluffadvisory.com, a minimal ad server on the essay pages, session-scoped synaptic audiences, a machine gate on every commit plus a human gate at a $0 floor, and public delivery numbers — house fill included.

  • Production live
  • AdCP media_buy · MCP
  • Cloudflare Workers · D1
  • Synaptic audiences
  • Dual governance gates
  • Cookieless · aggregate-only
  • 14-step live smoke

Brief Curated product Machine gate Human gate Serve Delivery report

Read the full breakdown → Open the seller agent ↗ Buy from it — the buyer’s guide →
02 Ask destination · AI visibility, run on itself

Yevgeny.ai.

A citation-first "ask the writing" destination over the published No Fluff corpus — one page, grounded answers with citations, live at www.yevgeny.ai. Presentation only: the brain is the shared ask-evgeny worker, a governed evidence system where a position graph rules retrieval, synthesis, and refusal.

  • Production live
  • Astro static
  • Cloudflare Pages
  • claude-sonnet-5
  • bge-m3 embeddings
  • Hybrid RRF retrieval
  • Position Records
  • Citation-first
  • Voice input

Question /api/ask Hybrid retrieval Unlisted filter Cited answer

Yevgeny.ai in miniature: a typed question flows through the cache check, retrieval over the corpus, generation, and the unlisted filter into a grounded answer whose citations trace back to the writing. Yevgeny.ai One page. The whole corpus, answering. The shared ask-evgeny worker is live at ask.nofluffadvisory.com — the Yevgeny.ai landing and the site widget speak the same /api/ask contract. LIVE WORKER The page sends the question with prior-turns-only history and deliberately no page field. A bare question — empty history, no page — is the only shape the worker's KV answer cache serves, so first questions and the suggested chips hit it. ask the writing KV answer cache — serves bare questions only. Entries live 7 days and the key includes the corpus version, so a contentful corpus sync invalidates every cached answer at once; enriched answers skip the cache in both directions. KV cache · 7d The published corpus — essays, glossary, pages, polls — indexed with bge-m3 embeddings; every count on the page is computed at build time, never hardcoded. Hybrid retrieval pulls the top 12 passages; unlisted passages are capped to a third of that and enter the prompt only as background that must never be cited, named, or quoted. The existing ask-evgeny worker — same /api/ask contract, grounding rules, and citation mapping as the site widget. Generation Anthropic claude-sonnet-5, embeddings Workers AI bge-m3. ASK-EVGENY WORKER one shared brain claude-sonnet-5 · bge-m3 Server-side unlisted-leak filter — checked after generation, so unlisted content can never appear in answers or citations. Prompt rules are the first layer; this is the hard one. UNLISTED FILTER The blocked branch — any sentence carrying an unlisted marker is stripped here before the answer ships; only the grounded remainder passes the gate. Grounded answers only: every claim comes from retrieved passages and carries a numbered citation — [S#] tags are mapped server-side to [n] markers; positions, numbers, and opinions are never invented. When the corpus doesn't cover it, it says so. GROUNDED ANSWER [1] [2] Every [n] marker resolves to a numbered source card — number, url, title, kind, excerpt — linking back to the published writing it cites. [1] essay [2] glossary Citations trace home: each numbered chip links back to the published passage retrieval surfaced — never to unlisted background, which must never be cited, named, quoted, or linked. the published corpus · counts computed at build time a bare question in — a grounded, cited answer out; when the corpus doesn't cover it, it says so
Read the full breakdown → Open Yevgeny.ai ↗
03 Editorial ops · A2A in production

The Editorial Agent.

An editorial agency run as agents, end to end — demand sensing from GA4 and Search Console, primary-source research, multi-agent production, an adversarial verify gate, CI publishing, creative distribution, and self-graded measurement. A human editor approves and owns the byline.

  • Live · weekly cadence
  • GitHub Actions
  • GA4 + GSC
  • Claude agents
  • Adversarial verify
  • Self-graded targets

Sense Brief Research Produce Verify Publish Measure

The Editorial Agent — one loop, eight handoffs, a rulebook and a human at the center THE EDITORIAL AGENT One loop, eight handoffs. RUNS WEEKLY · SELF-GRADED SENSE GA4 + GSC weekly pulse PLAN demand becomes briefs RESEARCH primary sources, cited PRODUCE pages + diagrams VERIFY the adversarial gate PUBLISH CI to the edge DISTRIBUTE carousels + video loops MEASURE snapshots, self-graded VERSIONED RULES the rubric agents answer to HUMAN EDITOR approves · owns the byline Demand becomes a brief; a brief becomes a page; the gate decides; the loop measures itself.
Read the full breakdown →
04 Agentic CDP · Built in public

AI Growth Operator.

An Agentic CDP for SMB brands — one operator running three deterministic recipes, with native holdouts and versioned operating rules. Two demo verticals, built in public with the spec in the repo.

  • Agentic CDP
  • One operator
  • Three recipes
  • Native holdouts
  • Versioned rules
  • Two demo verticals
  • Built in public

Brand events Operator Deterministic recipe Holdout split Measured lift

AI Growth Operator in miniature: the six-stage loop around one operator core, one human gate before money moves, and the three measurement labels every report must carry. AI GROWTH OPERATOR Six stages, one gate, three honest labels. 10% holdout Scan accounts, spend and signals for the next dollar worth moving. Find money The system drafts the plan, the audience and the creative — nothing ships yet. Draft The one human gate: money moves only after a person signs off. HUMAN SIGN-OFF Approve Approved work goes live across channels — no hand-offs. Activate Results are read against a 10% holdout that never sees the campaign. Measure Holdout deltas feed back in — the next loop starts smarter. Learn An Agentic CDP — one operator running three deterministic recipes, with native holdouts and versioned operating rules. AGO one operator agentic CDP · built in public EVERY REPORT CARRIES EXACTLY ONE A randomized holdout where exclusion is enforceable — owned-channel flows, audience of 500 or more. holdout-verified No control group available — the readout says so instead of claiming lift. before/after · no control group Where impact cannot be proven — Meta audience sync is always directional — no lift claim is made. directional one operator · three recipes · native holdouts — the loop pauses once, at the gate
Read the full breakdown → Launch the live demo ↗
05 Live prototype · Agentic orchestration

AdCP Ecosystem.

A live visual prototype for agentic advertising orchestration — exploring how signals, identity, governance, creative, buying, measurement, and clean-room workflows could coordinate as connected agents around a live brief and execution cycle. Experimental, not a production product claim.

  • Live prototype
  • AdCP
  • Agentic orchestration
  • Signals
  • Governance
  • Clean rooms
  • Measurement

Brief Agent fanout Governed actions Business trace Feedback

The AdCP ecosystem: connected agent roles — signals, identity, governance, creative, buying, measurement, clean room — orbiting a live brief, pulses relaying between them, a business trace feeding back. ADCP ECOSYSTEM Connected agents, one live brief. LIVE PROTOTYPE business trace Live brief context — connected agents operate around a live brief, execution cycle, and business trace. AdCP live brief Payload / JSON trace — every message inspectable, for transparency. { json trace } Signal fanout — intent distributed out across the connected agents. SIGNALS fanout Identity paths — what data an agent used, made visible. IDENTITY paths Governance paths — what rules applied to each agent action. GOVERNANCE paths The creative loop — one side of the execution cycle around the brief. CREATIVE loop The buying loop — where the brief becomes agent-driven execution. BUYING loop Measurement flow — what business signal changed, fed back as the trace. MEASUREMENT flow Clean-room flow — governed data collaboration inside the same orchestration. CLEAN ROOM flow Agentic advertising needs visible orchestration.
Read the breakdown → Open live ecosystem ↗
06 Founder · Operator

Signal-Stack.

AI-powered pre-call sales intelligence and a LinkedIn Network Analyzer, both running on Cloudflare's edge — built to turn relationship data into usable meeting intelligence.

  • Production live
  • Claude API
  • Cloudflare Workers
  • D1 / SQLite
  • Gmail · Gong · LinkedIn · CRM
  • Stripe
  • Privacy-by-design

Calendar Connectors Signal Computation Claude Synthesis Pre-call Brief

The pre-call brief pipeline: five sources pulse into signal computation, Claude synthesizes on the org's template, and the brief assembles line by line. THE PIPELINE Every signal in. One brief out. Gong call history — a 90-day lookback per prospect. Gong calls · 90-day lookback Email threads — the last 60 days of analysis. Gmail / Outlook threads · 60-day Profile signals pulled from LinkedIn. LinkedIn profile signals Deal and account context from Salesforce or HubSpot. Salesforce / HubSpot CRM context Contact enrichment via Clay. Clay enrichment The signal layer scores deal health, relationship velocity, champion stability, and deal-velocity trend. SIGNAL COMPUTATION Deal health Relationship velocity Champion stability Deal-velocity trend Claude synthesizes the signals into a structured brief using the org's own prompt template. CLAUDE synthesis org prompt template The pre-call brief lands by email or Slack the morning of every meeting on the calendar. PRE-CALL BRIEF delivered · email / Slack Every channel the deal touches, scored and synthesized into the morning brief.
Read the full breakdown →
07 Standards contribution · AdCP Signals & Measurement WG

AdCP Signals Adaptor.

A live, MIT-licensed reference seller implementation for agentic advertising — built to help buyer agents discover, query, activate, and reason over advertising signals through AdCP.

  • Production demo
  • MIT
  • AdCP 3.1.0
  • Cloudflare Workers
  • TypeScript
  • MCP
  • Discovery JSON

Buyer Agent MCP Request Schema Validation Signal Catalog AdCP Response

The handshake: a buyer agent's natural-language brief enters through the MCP socket; the adaptor validates, looks up and translates; spec-compliant signals stream into the seller catalog. THE HANDSHAKE A brief goes in. Spec-compliant signals come out. A buyer-side agent issues an MCP request — any AdCP-enabled buyer agent can call the adaptor. Buyer agent issues an MCP request The natural-language brief — query_signals_nl translates buyer intent into a structured signal lookup, no agent prompt-engineering required. BRIEF Built on Anthropic's Model Context Protocol — the socket every AdCP-enabled agent plugs into. MCP The reference seller implementation: every payload schema-validated against the vendored AdCP corpus, response header pinned to the released spec patch. Cloudflare Workers, MIT-licensed. AdCP 3.1.0 AdCP Signals Adaptor reference implementation validate · spec corpus look up · signals translate · MCP → AdCP Cloudflare Workers · MIT The seller catalog — signals across 14 verticals and 5 signal types; activate_signal turns a discovered signal into bookable inventory. SELLER CATALOG discovered get_signals activated activate_signal → bookable spec-compliant X-AdCP-Spec-Version pinned 14 verticals · 5 signal types One spec, many platforms. The bridges were the bottleneck.
Read the full breakdown →
What this proves

Four operator beliefs.
These systems live them.

  1. 01

    Standards need working implementations.

    Protocols become real when sellers and buyer agents can test them against running code — not slide decks.

  2. 02

    AI products need workflow gravity.

    Agentic systems only matter when they land inside the user's daily work. The brief shows up where the meeting lives.

  3. 03

    Edge-native stacks are enough.

    Many commercial AI products can ship without heavy infrastructure when the architecture is tight from day one.

  4. 04

    Advisory improves when it is built.

    Building exposes the real trade-offs behind the strategy. The thesis gets sharper when you have to ship it.

Advisory bridge

Want this kind of operator judgment applied to your company?

The builds prove the thesis. The advisory applies it to GTM, data, product, and operating-model decisions.

See advisory services Take the readiness scorecard
How this connects to advisory

Built systems route to specific playbooks.

Building proves the theses. Each system maps directly to the advisory work it informs — strategy tested against shipped code.

  1. 01 Standards work

    AdCP Signals Adaptor →

    Reference seller implementation for the agentic-advertising standard. Informs the advisory work where buyer agents, attention economics, and category creation matter most.

    Routes into
    See the full breakdown →
  2. 02 Product work

    Signal-Stack →

    AI-powered pre-call sales intelligence and a LinkedIn network analyzer running on Cloudflare's edge. Same orchestration pattern the advisory work argues for at the strategy layer.

    Routes into
    See the full breakdown →

Building is proof that the advisory is tested in shipped systems — not theory. The buyer journey routes both ways: advisory sharpens the build; the build proves the advisory.

Next step

Want this applied to your GTM, data, or product system?

The builds show how I think. The advisory applies the same operator logic to live commercial, technical, and organizational decisions.

See advisory services Take the readiness scorecard