Roadmap¶
Two things come before everything else: finish the whitepaper, and make the professional profile match the work. Both are about turning evidence that already exists into claims someone can find. Building more before doing that would be adding to a body of work nobody has been told about.
Phase 1 — Publish and position Sep – Oct 2026
Turn work that already exists into claims that can be found. No new building.
Whitepaper — sourcing pass
Done. Of 55 references, roughly two dozen were SEO articles, forum posts or aggregator pages — and four claims were substantively wrong, not merely weakly sourced. Four numbers deleted rather than re-sourced. The paper is shorter and considerably more defensible.
Whitepaper — original contribution
Rebuilt around a claim the evidence actually supports: you build the evaluation capability before you can define the metrics, and its first job is consistency rather than correctness. Sections 1–3 drafted, including the finding that three of four specialised retrieval paths were not behaving as the architecture described — and the answers still looked fine. Sections 4–7 outlined.
Whitepaper — editorial review, then publish
External review before release, not after. Publication on this site first, as the canonical version with a stable URL.
CV refresh to September 2026
Done. The CV ended at Sep 2025 while the Work section published a May–Sep 2026 case study — a twelve-month hole visible to anyone reading both. The 2026 engagement is now on the CV, sanitised to the same standard as the case study. This unblocks the profile rewrite.
Deploy the site
Done 2026-09-09. Live at probatio.yevgen.ponomarenko.io. Until that afternoon the site had never been deployed and the README claimed otherwise — everything published existed only locally and in GitLab, which meant the distribution strategy rested on a canonical record nobody could open. Applying it surfaced three faults that had been latent: an invalid root Terragrunt config that failed init, a globally-colliding S3 bucket name, and a CloudFront redirect that would have hidden the home page behind the CV.
LinkedIn profile alignment
Fully unblocked — the CV is current and the site is live, so the Featured links and the About URL work now. Rewritten from the CV: A positioning statement in the headline rather than a keyword string; the keywords go in the body, where search still finds them. This is manual work in the LinkedIn UI: copy is drafted here, applied there.
First expertise-claim post
The whitepaper launch, framed for a business audience, linking to the canonical version here. This is the first public claim to semantic architecture as a specialism, so it goes out after review, not before.
Phase 2 — Make the claims inspectable Nov – Dec 2026
A technical reader checks whether anything can be opened. Right now one thing can.
Semantic Metric Explorer
The first purpose-built app. No model calls, no secrets, no backend — a synthetic CPG metric model and a static front end, which is why it can ship rather than wait behind operational work. Makes the argument the case study already states into something clickable.
Public GitHub presence
A profile README anchored on the memoir platform, plus two seeded repositories: the confidentiality-guard pattern and the case-study template. This site's own source cannot be public, so the workshop lives elsewhere.
Second case study
The internal pre-sales review platform — an eight-dimension analysis pipeline over deal material. Needs an IP check and a sanitisation pass before anything is written; an employer gets the same disclosure discipline as a client.
Phase 3 — Compound Q1 2027
Work that only pays off once the first two phases exist.
Retrieval evaluation workbench
Vector, hybrid and graph retrieval compared on quality, citation coverage, precision, latency and cost — over a fixed public corpus, with the retrieved context inspectable. Pre-computed results first; interactive only once cost caps and a kill switch exist.
Zone Ontology
Built with my son. The same explorer engine on a completely different domain, which is the point: one engine, two datasets is a stronger claim that the pattern generalises than either alone.
Machine-readable index
A JSON index of work, methods, lab and writing. Discovery increasingly runs through systems that retrieve and cite rather than browse, which makes machine-legibility a distribution channel rather than a nicety.
Feed, then a newsletter
RSS is blocked — the site generator ships no feed plugin at this version, and faking one is worse than waiting. A monthly synthesis follows once there are enough substantive pages to synthesise.
Not committed no date
Listed so it is clear these were considered and deliberately left un-scheduled.
Artifact pipeline
Scheduled sync of professional artefacts, an API over them, and LLM analysis flagging where the CV, this site and LinkedIn have drifted apart. A good idea that solves a problem I do not yet have at a scale that justifies it.
Decision-log generator
Real personal utility, no urgency. It moves up if the manual decision records here become tedious enough to automate.
A chatbot over this site
Deliberately never. Core professional material does not go behind a chat interface. Navigation and source pages come first, and if they are good the assistant adds little.
What already shipped¶
Done Sep 2026
Recorded because a roadmap with no completed column is a wish list.
Site restructured around five layers
Position, proof, method, inquiry, contribution. Three placeholder pages deleted rather than filled.
First case study published
The RGM knowledge platform, with its maturity stated as a prototype and what was not achieved recorded on the page.
Four method pages
Fail-loud pipelines, intent routing, competency-question evaluation, evidence gates — each with the conditions under which it does not apply.
Confidentiality firewall in CI
Negative-tested before being trusted. Caught a real incident before it became a publication.
First publicly inspectable artefact
The memoir publishing platform — public source, four languages, three output formats.
What depends on what¶
The order above is not preference. Four dependencies actually constrain it:
graph LR
CV["CV refreshed<br/>to Sep 2026"] --> LI["LinkedIn profile<br/>alignment"]
WP1["Whitepaper<br/>sourcing pass"] --> WP2["Original<br/>contribution"]
WP2 --> WP3["Editorial review<br/>and publication"]
WP3 --> POST["First expertise-<br/>claim post"]
LI --> POST
APP["Semantic Metric<br/>Explorer"] --> GH["Public GitHub<br/>presence"]
GH --> HN["Show HN"]
APP --> ZO["Zone Ontology<br/>shares the engine"]Stated plainly, in case the diagram does not render:
| This | must happen before | this | because |
|---|---|---|---|
| CV refreshed | → | LinkedIn profile rewrite | The profile is a rendering of the CV, not a separate document. Rewriting it first guarantees they diverge. |
| Whitepaper sourced and reviewed | → | The first expertise-claim post | The post is the claim; the whitepaper is the evidence. Publishing the claim first is the thing this whole site is built to avoid. |
| Semantic Metric Explorer | → | Public GitHub presence | An empty profile is worse than no profile. The repository needs something in it. |
| Semantic Metric Explorer | → | Zone Ontology | They share an engine. Building it in the professional context first means the second one costs a weekend. |
The rule underneath all of it: never post without an artefact behind it. Distribution follows work; it does not substitute for it.
The whitepaper track, in detail¶
The first Semantic Architecture claim, so it gets the most care.
- Sourcing pass. Every load-bearing claim traced to primary literature, product documentation, or a measurement I took. Where a claim rests on someone else's summary, either replace it or mark it as such. This is the slow part.
- Original contribution. Drawn from the platform build. The claim: you build the evaluation capability before you can define the metrics, because the instrument is what tells you which metrics matter — and its first job is consistency, not correctness. Evidence is three instruments that each catch what the others cannot see, and the route-targeting finding that three of four specialised paths diverged from the architecture while the answers still looked fine. Sanitised the same way the case study was.
- Editorial review. External, before release. The reviewer's job is to find the paragraphs that sound right and are not.
- Publish here first, with a stable URL, a
last verifieddate and a status. This is the canonical version. - Then the post, written for a business reader, linking here. Not a summary of the paper — the one idea from it that changes what a reader does on Monday.
The LinkedIn plan, in detail¶
LinkedIn is a doorway, not an archive. Everything canonical lives here; the profile is a rendering, and posts are entry points.
Profile, in order:
- Headline — the positioning statement, not a keyword string. "Data & AI Leader | Power BI | Databricks | GenAI" is searchable and forgettable; the keywords belong in the About and experience text, where search still indexes them.
- About — the 150-word narrative from the About page, which is already written and already true.
- Experience — aligned to the refreshed CV. Same facts, same dates, same contribution scope. Divergence between the two is the thing a hiring manager notices.
- Featured — the case study, the whitepaper once published, and the memoir platform repository. Three items, each inspectable.
Cadence: roughly two posts a month, each tied to something that exists. A post per artefact, never an artefact-free post. Three formats work:
| Format | Source | What it does |
|---|---|---|
| Launch | A published whitepaper or app | Makes a claim, links to the evidence |
| Distillation | A method page | One idea, argued in 200 words, for readers who will never open a method page |
| Field note | Something that failed or changed my mind | The most useful and the least common on that platform |
What is measured: nothing yet, honestly. There is no analytics on this site, so the loop from post to visit to conversation is currently invisible. Deciding what to measure — and whether it is worth measuring at all for a profile this size — comes before adding tracking.
How this page is built: hand-written HTML and CSS, no JavaScript, so it renders for a crawler, prints correctly, and cannot break. The one diagram is Mermaid, with the same information in a table beneath it. Reasoning in the site's tooling decision record.