SLAVA SOLODKIY / Founder · Systems Builder
FINTECH FOUNDER · AI SYSTEMS BUILDER · NVIDIA FIRST PLACE / LONDON 2026

solodkiy.cv
TECH

I’m a fintech founder, product architect and AI-native systems builder. I design and direct applied AI systems across hardware, software, institutional trust and product constraints — with a practical commitment to NVIDIA + AMD + Intel portability, not single-vendor lock-in.

BUILT IT.
TESTED IT.
SHIPPED IT.

NVIDIA / #1
AGENTIC AI
LONDON 2026

OPEN THE EVIDENCE ↗
#1FIRST PLACE
06 OCT 2026
NVIDIA CLAW AGENT CHALLENGE / LONDON / EXTERNAL RECOGNITION

An agent that cannot authorise itself.

AuthorityClaw — a working local-AI payments agent. The model reads invoices; a deterministic gate checks delegated authority before a simulated payment executes. Won first place and was invited to demonstrate at NVIDIA's Build-a-Claw London on 14 October.

4/4 E2E TESTS
6 INVOICES · 7 RECEIPTS
LIVE REVOCATION
SEE THE WINNING BUILD ↗
Local / Sovereign AI Agentic Systems Formal Methods Digital Identity AI Evals Regulated Systems AI Infrastructure Vibe Coding → Evidence
FOUNDER-LED SYSTEMS ARCHITECTURE / PROOF THROUGH EXPERIENCE

What I have built, tested & learned.

I do not measure technical ability by whether I memorised the syntax. I measure it by whether I can frame the problem, specify constraints, direct the build, inspect the evidence, catch bad results and turn both results and failures into reusable systems and the next research question.

0–4
numbered lab nodes / experience
Olares One, Minisforum, Morefine, Acemagic and NVIDIA DGX Spark.
3
compute families in scope
NVIDIA / CUDA · AMD / ROCm · Intel / Arc/XPU research.
128
tests in a full-stack identity reference build
AI-assisted implementation; inspectable repo and test suite.
0
AI authority override in the decision-plane demo
The model can propose evidence. It cannot overrule controls.
01 / LOCAL INFERENCE

Deploy & compare local models

GGUF / llama.cpp / Ollama-style runtimes, quantisation choices, context, VRAM, latency and practical fit across NVIDIA and AMD hardware.

02 / AGENTS

Orchestrate cloud → local execution

Use cloud reasoning as the planning/review layer while Linux-hosted CLI agents, local models and services do the execution.

03 / EVALS

Turn “seems better” into a test

Baselines, matched prompts, controlled conditions, synthetic cases, regression checks, raw evidence and explicit failure criteria.

04 / FORMAL METHODS

Separate AI output from authority

Deterministic validation → First-Order Logic / Prolog → institutional policy / OPA, with AI kept outside the final authority path.

05 / MULTIMODAL

Run speech, image & research stacks

Local STT/TTS, image generation/editing, research pipelines and specialist model roles rather than forcing one LLM to do everything.

06 / PRODUCT SYSTEMS

Direct AI-assisted full-stack builds

APIs, data models, onboarding logic, web/mobile surfaces, test suites and public technical artefacts — without pretending I hand-coded every layer.

FIVE-NODE EXPERIENCE
PLATE 01

Five nodes. Several architectures. One learning curve.

The five numbered machines document equal parts of my professional, hands-on experience with inference, operating systems, hardware interoperability, memory constraints, agent interfaces and recovery. Each contributes a distinct architectural lesson to the same body of work.

NODE 0 / OLARES ONE · NVIDIA CUDA · SYSTEMS & UX

Olares One

96 GB RAM · RTX 5090 Laptop 24 GB · OlaresOS / Linux + Windows VM
  • Local AI, Nemotron–FOL and controlled institutional decision experiments.
  • FreeToken / large-MoE expert offload beyond the simple VRAM-size rule.
  • Browser-based Olares management, agent-driven administration and the boundary between web UI and native CLI.
  • Image, speech and research workflows, alongside remote operator and recovery experiments.
NODE 1 / MINISFORUM · AMD ROCm · CROSS-VENDOR TRANSITION

Minisforum M1 Lite

96 GB RAM · Radeon AI PRO R9700 32 GB · OCuLink · Linux/OlaresOS + Windows VM
  • AMD inference from earlier RDNA3/RX 7900 XTX and Vulkan/Windows investigations to the R9700 ROCm/HIP path.
  • Qwen GGUF latency, quantisation and VRAM-fit comparisons; specialist Heretic 27B backend.
  • AnythingLLM multi-model HQ, Kubernetes/Ollama and Windows-to-Linux API bridges.
  • CUDA-to-HIP/ROCm kernel-porting research: CPU fixture and draft HIP port recorded; GPU execution not yet demonstrated in that sprint.
NODE 2 / MOREFINE · SMALL-GPU ECONOMICS · SPECIALIST SERVICES

Morefine M900

48 GB RAM · RTX 3060 12 GB · OlaresOS/Linux + Windows VM
  • FreeToken hybrid expert offload with a 35B-class MoE model despite 12 GB GPU VRAM in a documented configuration.
  • Whisper speech-to-text, local TTS and dedicated utility workloads.
  • Earlier Nemotron/router and GGUF runtime experiments; later abliterated Qwen 9B via llama.cpp.
  • Windows desktop, local inference APIs, Kimi-assisted cross-node orchestration and readback.
NODE 3 / ACEMAGIC · MULTI-AGENT EXECUTION · RETRIEVAL

Acemagic AM18

32 GB RAM · RTX 3060 12 GB · OlaresOS/Linux + Windows VM
  • Local model serving from Nemotron experiments to a specialist Dolphin 8B GGUF backend.
  • SearXNG, reranking, Crawl4AI-style research components and the limits of chaining them into real browser tasks.
  • Intel Arc B580 12 GB — tested 10 October 2026. A documented OCuLink experiment validated the OlaresOS → xe → Kubernetes → OpenVINO path with bge-reranker-v2-m3. Both /v1/models and /v1/rerank returned HTTP 200, confirming a working inference service. Recorded experiment ↗
  • Claude Code/CLI on a Windows VM calling specialist Linux inference through a shared AnythingLLM HQ.
  • Remote terminals, host/guest separation, API integration, persistence tests and recorded recovery boundaries.
NODE 4 / DGX SPARK FOUNDERS EDITION · NVIDIA GB10 · UNIFIED MEMORY

NVIDIA DGX Spark — one machine, two operating systems.

GB10 Grace Blackwell · ARM64 · 128 GB unified memory · DGX OS and OlaresOS
  • Benchmarked local 120B inference on the factory DGX OS; worked with CUDA, NVIDIA Sync, image generation and long-context experiments.
  • Reinstalled the machine with OlaresOS, diagnosed Secure Boot / kernel module / NVML / HAMi failures and restored GPU availability.
  • Compared the same hardware through two different software and agent-operation experiences rather than declaring a benchmark winner.
  • Published the technical field study and reproducibility evidence, including caveats and follow-up questions.
TWO DEDICATED AI APPLIANCES / DIFFERENT MEMORY DOMAINS

Tiiny Black + Tiiny White

Two independent NPU devices, each with a vendor-described 32 GB SoC memory pool and 48 GB dNPU memory pool. Black provided Qwen3.6 35B-A3B; White provided GPT-OSS-120B. Both were integrated as distinct AnythingLLM HQ endpoints, with inference and chat-history evidence.

The open research question is how SoC/dNPU memory placement changes actual useful work, not whether “80 GB” looks good on a specification sheet.

Original heterogeneous compute + Tiiny questions ↗
SYSTEM ARCHITECTURE / SHARED CONTROL PLANE

Heterogeneous inference, one decision boundary.

Specialist models on AMD, NVIDIA and NPU endpoints connect through a shared workspace. The aim is to move perception and compute without moving authority: human mandates, deterministic checks and provenance remain separate from probabilistic output.

Five model workspaces and history readbacks were documented. That does not yet imply automatic routing, durable operations after all reboots or a financially self-sustaining autonomous organisation.

AuthorityClaw · separation of model and mandate ↗
HETEROGENEOUS BY DESIGN
RESEARCH THREAD 2026

Not another CUDA-only future.

NVIDIA is a powerful reference platform, but I deliberately work across NVIDIA/CUDA, AMD/ROCm and Intel Arc/XPU as a portability research direction. The point is not to favour a logo: it is to learn which parts of the workload, memory and authority model can move between architectures — and where vendor-specific dependencies really bite.

01 / PORTABLE COMPUTE

CUDA → HIP/ROCm → XPU

Test the same useful pipeline across hardware and runtimes. Record API gaps and control-plane differences. A November CUDA-to-HIP softmax micro-sprint reached a CPU reference and a draft port, not GPU performance parity; that remains a precise open result.

Intel/Arc portability research agenda ↗
02 / MEMORY AS ARCHITECTURE

VRAM, RAM, unified memory, NPU pools & SSD.

Different memory arrangements change what can run, and at what latency or power. Experiments span FreeToken expert offloading, AMD dedicated VRAM, GB10 unified memory and Tiiny split SoC/dNPU pools. WiCi InferSSD, Strata and similar approaches are research targets, not claimed as fully validated on every device.

FreeToken measurements ↗
03 / PROBLEM → EXPERIMENT → NEXT QUESTION

Every result opens the next brief.

My field notes do not stop at a PASS or a token-per-second chart. They preserve failures, falsifiers, unexplored dependencies, future tests and the practical human problem that would make the work worth repeating.

Six future research questions ↗

The research boundary is intentional: running a proven model, drafting a port, verifying GPU equivalence and demonstrating commercial reuse are four different results.

FLAGSHIP TECHNICAL CASE
PLATE 02

AI can propose. It does not get authority.

My most useful technical thread so far: a reproducible synthetic demonstrator for regulated decisions, where model output must survive deterministic and formal checks before institutional policy can act.

SOVEREIGN DECISION PLANE / LIVE CASE 01

Same decision plane.
Different evidence.
Different outcome.

A local demonstrator that treats the LLM as a perception layer — not as the final decision authority. The public repo includes a reproducible path, committed artefacts and CI.

“The model may perceive and explain. It cannot override a failed deterministic control.”
DECISION PATH / MACHINE-INSPECTABLE
01 / Evidence
02 / Nemotron candidate observations
03 / Deterministic verification
04 / SWI-Prolog / First-Order Logic
05 / OPA / Rego institutional policy
06 / ALLOW · ESCALATE · BLOCK
07 / Evidence packet + SHA-256
GREEN / CONTROL
FORMAL: accept_extraction
ACTION: ALLOW
AI OVERRIDE: false
RED / MODEL-DERIVED
FORMAL: request_better_evidence
ACTION: ESCALATE_TO_HUMAN
AI OVERRIDE: false
SELECTED BUILDS / 2026
PLATE 03

Things you can open.

Not a gallery of every page I have made. A short evidence set showing the range: infrastructure, evaluation, identity systems, formal reasoning and AI-assisted product builds.

SLAVAOS / SYSTEMS ARCHITECTURELIVE LAB · NOV 2026

Building an AI workforce before AI workforces actually work.

A founder-led architecture and learning programme spanning the five numbered Olares, AMD/NVIDIA and DGX Spark machines, standalone Tiiny appliances, operating systems, AI runtimes, AnythingLLM workspaces and governed agents. My role is to define the problem, architecture, acceptance criteria and failure boundaries and direct AI-assisted implementation — not to pretend I hand-coded every component.

Why: to make AI tools reusable for real work and test whether a personal agent infrastructure can eventually become economically productive. The key obstacle is not token speed; it is reliable hand-off, persistent state, remote access and authority.

VERIFIED IN PROJECT REPORTS / FIVE SPECIALIST HQ WORKSPACES + PERSISTENT HISTORY; WINDOWS VM → LINUX BRIDGES ON N1/N2/N3. NOT YET PROVEN / REBOOT-DURABLE FULL-FLEET OPERATIONS, GENERAL TASK QUALITY OR SELF-SUSTAINING COMMERCIAL EXECUTION.
FEATURED / AUTHORITYCLAWNVIDIA · FIRST PLACE · 2026

The agent cannot authorise itself.

A local Qwen model reads synthetic supplier invoices. A separate deterministic authority gate checks principal, capacity, mandate, scope and revocation. Only a valid decision receipt can reach the simulated payment rail — including under bank-detail changes and prompt-injection attempts.

PROOF / 4 OF 4 E2E TESTS · 6 INVOICES · 7 HASHED RECEIPTS · VERIFIED CHAIN · NO REAL MONEY MOVED
AUTHORITY / DECISION PATH
01 / MODEL READS THE INVOICE
02 / AUTHORITY GATE DECIDES
03 / ALLOW · HOLD · REJECT
04 / VALID RECEIPT → SIMULATED RAIL
AI-01 / LOCAL INFRASTRUCTURELIVE LAB

Heterogeneous inference & memory experiments

Hands-on investigation of heterogeneous inference across NVIDIA DGX Spark (GB10), Olares One, AMD Radeon and NVIDIA GPUs, plus Tiiny NPU appliances. I compare memory architecture, model offloading and practical usefulness — not merely token rates.

PROOF SIGNAL / DGX Spark, Olares One & cross-vendor experiments; R9700 tests; five model routes
ID-02 / FULL-STACK REFERENCE BUILDPUBLIC REPO

LarePass Reference System

AI-assisted self-sovereign identity reference system with web + mobile surfaces, PostgreSQL, OpenAPI 3.1, generated clients and a branching KYC/KYB onboarding engine.

PROOF SIGNAL / 40+ endpoints · 36-node flow · 128-test suite
FL-03 / FORMAL REASONINGEXPERIMENT SERIES

FOL × Qwen × Nemotron

Experiments around whether natural-language model outputs can be translated into explicit, inspectable logic and then constrained by deterministic or institutional policy layers.

PROOF SIGNAL / Prolog · OPA/Rego · evidence regeneration · reproducible demo
RG-04 / REGULATED SYSTEMSDESIGN + PROTOTYPE

Revolut / Governed Agent

A set of experiments asking a harder question than “can an agent act?”: under what evidence, mandate, scope and revocation state should an agent be allowed to act?

PROOF SIGNAL / authority schema · scope · provenance · revocation · deterministic gates
MM-05 / MULTIMODAL LOCAL AIWORKING TOOLS

Audio + Image Local Stacks

Local speech recognition and TTS on a dedicated NVIDIA worker; ComfyUI image generation and editing on the main node; specialist workloads kept separate from the main LLM path.

PROOF SIGNAL / local WAV · local transcription · generated & edited images · isolated runtimes
DI-06 / DOMAIN DEPTH2025—2026

Digital Identity Research

The domain behind much of the technical work: reusable identity, compliance-grade architecture, proof-of-personhood, government wallets, KYC/KYB and AI-era institutional trust.

PROOF SIGNAL / public research repository · releases · reports · DOI-linked work
AI-NATIVE BUILD MODE
PLATE 04

I don't compete with programmers at typing code.

My useful skill is the loop around the code: turning a hard problem into an executable experiment, using specialist AI and tools as leverage, and refusing to accept a green screen without evidence.

01FrameWhat is the actual problem? What is not the problem?
02SpecifyConstraints, invariants, schemas, expected outcomes.
03DelegateChoose cloud agent, CLI, local model or specialist tool.
04BuildLet the machine do machine work; keep architecture inspectable.
05AttackCounterexamples, false premises, edge cases, matched comparisons.
06VerifyDeterministic checks, tests, formal logic, CI, raw evidence.
07Ship & ask againPublish evidence, limitations and the next problem worth solving.

Not claiming

  • career software engineer
  • ML researcher who designs model architectures
  • CUDA / kernel / compiler specialist
  • that every line in my repos was typed by me

Actually doing

  • AI-native technical product building
  • systems prototyping & orchestration
  • local inference and model evaluation
  • agent supervision and technical specification
  • formal decision experiments
  • AI-assisted full-stack prototyping
REVISION LEDGER / HOW EVIDENCE CHANGES THE SYSTEM
INCIDENT → INVARIANT → TEST
R001 / CHANNEL
A real channel is evidence, not authority. Government-domain origin can start verification; it cannot release data.
R002 / HOLD
Irreversible disclosure stays null until authority resolves. A “small provisional payload” still leaks. HOLD means zero outbound PII.
R003 / CAPACITY
Identity is not enough; current capacity must also hold. The ontology changed when the evidence showed that role and authority can diverge.
WORKING TOOLBOX
PLATE 05

The stack I can reason about and operate.

Not a keyword cloud. These are tools, layers and concepts that have appeared in working experiments or current builds.

LayerCurrent toolboxWhat I use it for
Local LLM Qwen · Nemotron · Mistral · Gemma-class experiments · abliterated sandbox models Inference, extraction, synthesis, coding, adversarial comparisons.
Runtime llama.cpp · Ollama · LM Studio · FreeToken experiments Local serving, GPU offload, quantised inference, heterogeneous hardware tests.
Agents ChatGPT/Claude / Codex/Caude CLI · desktop/CLI hand-off patterns Plan → execute → inspect → revise loops across cloud and local environments.
Formal SWI-Prolog · First-Order Logic · OPA / Rego · Z3-direction experiments Make assumptions explicit; keep policy and AI output separable.
Research Local Deep Research · SearXNG-style search layer · local LLM synthesis Evidence-led research without treating the base model as a current-information oracle.
Media ComfyUI · FLUX-class image models · Whisper · local TTS Local image generation/editing and speech workflows.
Product build OpenAPI · React / Vite · React Native / Expo · PostgreSQL · Docker Compose AI-assisted reference systems and prototypes with inspectable interfaces and tests.
Evidence Git · GitHub Actions · hashes · manifests · raw outputs · regression suites Turn a demo into something another person can inspect and rerun.
WHY THESE PROBLEMS?
PLATE 06

I entered tech through institutions, not computer science.

Banking, fintech, compliance and digital identity taught me to care about who is allowed to do what, on whose evidence, under which rules.

Not a career change — the same profession.
Only the subject of trust changed: first the institution, now the model.

That is why my technical experiments keep converging on authority, evidence, provenance, identity, local control and regulated decision-making.

I am not trying to become a better programmer than programmers. I am learning to become very good at turning difficult institutional problems into systems that can be built, attacked, measured and inspected.

// OPEN TO HARD PROBLEMS

Give me a messy system with real constraints.

I’m most useful where product, AI, regulation, identity, infrastructure and institutional reality collide — and where a slide deck is not enough. I like turning the problem into something we can run.