KS Kirill Shokhin

Independent research engineer

A working position across machine learning, formal methods, and the design of complex systems.

The structure is older than any vocabulary that names it
§ 01 Method

The same operation, in any vocabulary

Most of the work — from clinical decision support to formal protocols for hierarchical action, from rebuilding published SOTA to deep readings of emerging neural architectures — turns on the same operation: see the invariant first; treat any one domain's vocabulary as a projection of the invariant, not its definition.

A problem usually arrives wearing one domain's language. The first move is to refuse to stay in that language exclusively. The same task, restated in a second and a third vocabulary, exposes which parts are local to a phrasing and which parts survive translation. What survives is the object to work with. The rest is decoration that any one domain happens to add.

Operational form — three moves

  1. Define the object by its external arrows. Inputs, constraints, dependencies, sources of authority — including the top-level ones a commercial system always has and analyses usually omit.
  2. Find the decomposition where each arrow lands cleanly in one component. Test: if a component is modified, which external arrows change? Crisp answer — clean boundary. Diffuse answer — decompose again.
  3. Let the internal structure fall out. With interfaces fixed and orthogonal, implementation is constrained optimisation, not a design choice.

Each significant decision traces back to axioms or constraints. Alternatives are taken to the level of a usable artifact, then compared. Open questions are marked as external-dependent, not masked. The same procedure applies recursively to the code itself: module boundaries forced by the same external arrows, not by stylistic convention.

§ 02 Sources

Where the method was forged

Each item exists because it explains a property of how the method works now.

  • Theoretical physics — Landau–Lifshitz tradition. Source of the discipline: every claim traces to axioms; alternatives are computed, not vibed.
  • Production ML and computer vision. Pipelines deployed at broadcast level, alongside end-to-end independent work that outperformed a published SOTA from a major lab and went to production at the client. Source of production grip: structure has to survive deployment, not just whiteboards.
  • Architectural research. Mapping of the post-transformer design space — public technical analysis of emerging architectures like Mamba / state-space models, and unpublished hypotheses on tokenisation and in-model reasoning. Source of architectural fluency.
  • Multimodal DL technical leadership. Multi-year R&D-grade system taken from an open-ended question to clinical deployment under industrial constraints. Source of operating the method under live engineering load.
  • Independent formal work. A framework derived from minimal axioms to an operational protocol with impossibility results on the foundations. Source of formal closure on the method itself.
§ 03 Cases

Current instances of the position

Each starts in a different vocabulary and reaches the same operating frame. The trace, in every case: what was unclear how to do → what was seen → what was built → what came out.

Watermarking — replication, dissection, override

A published SOTA from a major lab, with the brief: outperform it. No off-the-shelf path; the model and the methodology were the whole problem.

Reimplementing the paper end-to-end exposed which architectural choices were jointly bounding its results — the published numbers were a ceiling imposed by those choices, not by the task. Those components were isolated and replaced; the new system outperformed the original across all four reported axes: detection accuracy, imperceptibility, robustness, and inference speed.

Deployed to production at the client. Results unpublished at client's request.

Formaesthetics — invariant across five vocabularies, taken to clinical deployment

A clinic asked for a tool. The task sat at the intersection of five vocabularies that no one had connected operationally — plastic surgery, perceptual aesthetics, 3D geometry, machine learning, and crowdsourced consensus. The actual question was whether objective planning of aesthetic surgery is possible at all.

The work was apparatus-first. The full chain — capture, unified topology, statistical shape space, ground truth, predictive model — had to be built before the original question could be properly posed. Four candidate representations were taken in parallel to working form and compared on independent criteria; the choice among them was forced by the data regime, not by inertia.

Technical leadership end-to-end: architecture, integration, and engineering quality were owned single-handedly; a small part-time team handled bounded algorithmic sub-tasks.

The whole system was then wrapped in a cryptographic structural-transparency layer that left no link of the chain editable after the fact — an audit guarantee with no precedent in clinical ML deployments.

Handed off to clinical operations at a Moscow plastic surgery clinic over a 20-month engagement; currently in surgical-team rollout. A research-grade technical report is the source for a forthcoming series of targeted papers.

GFSO — a protocol-level invariant for working relationships

Hierarchical work — handoffs, acceptance, escalation — still runs on something close to barter. Every organisation defines “valid” locally, claims compositionality, and discovers at the top level that local checks do not stack. This becomes acute the moment those systems start admitting AI agents alongside humans, and require an interface that holds regardless of which party sits on either end.

The framework is forced from two axioms — verifiability of goals, decomposability of complex tasks. Key constructions admit no alternatives; the protocol is derived, not designed. The result is agent-type-invariant by construction: the same rules hold whether the parties are humans, LLMs, or organisations, because the conditions are stated on the work, not on who does it.

Released open-source, with the engine implemented in pure Python. A public applied agent built on an earlier version of the framework carries benchmark traces on competitive-programming tasks and parts of Humanity's Last Exam — under a deliberately weak base model. The two are kept structurally orthogonal — the former is the load-bearing artifact, the latter is one of several possible applied surfaces.

§ 04 Artifacts

Public artifacts

GFSO — formal framework and engine github · open source Formaesthetics — clinical project, public page landing · public snapshot GFSO-agent — applied agent on an earlier version of the framework github · runnable benchmarks Mamba / state-space models — long-form deep-dive habr · ≈ 20k readers in the first week

in progressGFSO empirical preprint; the next applied agent on the current framework; two papers extracting methodology from the Formaesthetics work; a series of focused technical pieces on individual algorithms from the same source.

§ 05 Engagement

Selection by type of task

Independent. Not seeking staff positions. Selection is by type of task, not stack or domain.

relevantthe system is tangled and we don't know how to decompose it cleanly
notimplement feature X with stack Y

Project-scoped advisory, structural audit, R&D-heavy technical leadership.

§ 06 Contact

Write — RU or EN

mail kashokhin@gmail.com github github.com/Kirill-Shokhin linkedin linkedin.com/in/kashokhin telegram @KirShokhin