Andriy

ANDRIY CHOS
Founding AI Engineer • AI Systems & Infrastructure Architect

[open contact info](look above in the "contact info" section) telegram: @jizelfi Zhytomyr, Ukraine (Remote)
Por%olio: h*ps://github.com/40chos/oma-por%olio

PROFESSIONAL SUMMARY

Founding AI Engineer and Systems Architect with 3+ years shipping producKon AI infrastructure end-to-end — from self-hosted
GPU clusters and determinisKc orchestraKon engines to the agenKc systems running on top of them. Built KHASH, a company-
agnosKc, determinisKc-first automaKon architecture (formal 13-part design methodology, self-hosted Proxmox/vLLM
infrastructure, 187 catalogued connectors), and OMA, an autonomous Odoo ERP development agent built on it — plus 30+
other producKon AI systems integraKng Anthropic, OpenAI, Gemini, and self-hosted open-source models. DeterminisKc-first
by design: AI is invoked only where tested, human-approved logic cannot already decide the outcome, with human sign-off
gaKng real risk. Deep pracKcal experience with RAG architectures, MCP, mulK-agent orchestraKon, and hybrid Computer Vision
+ LLM pipelines — comfortable owning full delivery from infrastructure to producKon deployment.

TECHNICAL STACK
AI Coding Tools: Claude Code, Cursor, Windsurf, GitHub Copilot, Codeium
LLM APIs & Models: Anthropic
Claude, OpenAI, Gemini, Mistral,
Qwen, Local LLMs

Languages: Python, TypeScript, SQL, JavaScript, C++, C#, Go, Rust Frameworks: PyTorch, FastAPI,
Next.js, Node.js, PydanKc, LangGraph

RAG & Vector and Storager: ChromaDB, FAISS, Pinecone, Weaviate, DevOps & Infra:
OpenAI Embeddings, Sentence Transformers, PostgreSQL, Redis, RabbitMQ, Docker, Kubernetes, Proxmox, Linux
Neo4j, Qdrant (systemd), AWS, GitHub AcKons,
CI/CD, Keycloak, Teleport

AgenEc AI: MCP (Model Context Protocol), MulK-agent orchestraKon,
Tooluse, FuncKon calling, Autonomous pipelines, MulK-agent
build/review/QA systems, Governance-as-code, Cryptographic verificaKon

WORK EXPERIENCE

AI Systems Architect | BadkamerVakman B.V. (Dutch Construc5on & Renova5on Enterprise, 2,700+ projects)
Self-employed / Remote Contract • March 2026 – September 2026

▸ Engaged by BadkamerVakman B.V. (Netherlands, 2,700+ projects delivered) to automate internal operaKons end-
to-end — not with a single company-specific chatbot, but a durable, company-agnosKc system (KHASH) designed to outlast a
typical 5-year vendor-locked build and be instanKated for any mid-sized organizaKon, not hand-fit to one client.
▸ Own the full self-hosted infrastructure underneath it: a mulK-hypervisor Proxmox cluster (12 physical hosts, 91
VMs) running internal PKI (step-ca), Vault + Vaultwarden secrets management, Keycloak OIDC/SSO, Teleport audited SSH,
Postgres (pgvector), Redis, RabbitMQ, Qdrant/Neo4j for vector + graph retrieval, and Grafana/Loki observability — plus
Kubernetes-orchestrated workloads for services that needed container-level scaling rather than dedicated VMs.
▸ Run a self-hosted GPU model fleet (coding, general-purpose, embedding, reranking, vision, and speech models
served via vLLM) so the plalorm runs fully offline/sovereign rather than depending only on an external AI vendor — the
equivalent of a dedicated inference cluster most teams would otherwise rent from a cloud AI provider.
▸ Designed the architecture for reliability first: generated mulKple independent candidate designs, stress-tested each
against 16 real organizaKonal scenarios, resolved disagreements with targeted tests rather than opinion, and converged on
one fault-tolerant, determinisKc architecture — cryptographically frozen before development began so the foundaKon
couldn't silently driq.
▸ Formalized the converged design as a 13-part design-decision methodology: system context, capability modeling,
trust boundaries, interfaces/contracts, data & state, idenKty/PKI/secrets, network trust zones, execuKon & adapters,
AI/inference, sovereignty & offline operaKon, audit & evidence, deployment & lifecycle, and genesis & emergency-recovery
authority.
▸ Defined a formal integraKon taxonomy (8 roles × 7 kinds — Tool Adapter, Knowledge Connector, Infrastructure
Driver, IdenKty IntegraKon, Secret Provider, Event IntegraKon, Observability IntegraKon, and more) and applied it across 32
producKon services and 187 catalogued connectors spanning 12 categories (AI models, business systems, DevOps,
communicaKons, core infrastructure, knowledge/research).
▸ Built ~2,500 atomic automaKon primiKves across 50+ categories, many extracted directly from the client's own
historical customer emails — mined into recurring business processes and converted into determinisKc tool-plus-flow
definiKons: reviewed and approved once, then reused as building blocks rather than regenerated by AI each Kme. Combined
into self-verifying composite operaKons (e.g. a backup composite that independently checksums its own output) and higher-
level workflows on top.
▸ Proved the foundaKon supports building genuinely new AI systems, not just automaKng what already exists —
shipped OMA (Odoo Manager Agent), an autonomous system built directly on KHASH's capability model that turns plain-
English requests into real, installed Odoo ERP module code (models, fields, views, security rules).
▸ OMA's stack: custom hand-built orchestraKon (no LangChain/LangGraph) coordinaKng Manager, Build, Code-
Review, and TesKng/QA components; self-hosted Qwen models (27B reasoning, 30B coder, 9B fast-extracKon) served via vLLM
behind llama-swap on dedicated GPU workers, with Anthropic's Claude Sonnet as a cloud escalaKon Ker for the hardest calls;
FastAPI + PydanKc, PostgreSQL, a Neo4j knowledge graph grounding generaKon against the real live Odoo schema, Redis, and
naKve Odoo XML-RPC integraKon.

▸ Built the operator-facing interface in Next.js/TypeScript, that was surfacing real-Kme system state (91 VMs, GPU
fleet health) without overwhelming a human operator — i.e., designing informaKon hierarchy under high data velocity. Framed
as building lightweight, typed adapter services (Node.js/TypeScript) within the 187-connector integraKon layer, where the
hard problem was defining a consistent typed interface contract across wildly different upstream systems (business tools,
DevOps APIs, comms plalorms).
▸ Engineered determinisKc operaKonal rouKng engines from historical unstructured customer-email data,
automaKng client triage without generaKve hallucinaKons.

Founder & Lead AI Engineer | LegalTech Startup (U.S. Client)
Self-employed / Remote Contract • April 2025 – April 2026

▸ Own the full engineering lifecycle of a U.S. legal firm's AI infrastructure — from architecture decisions to producKon
deployment. ConKnuously design, iterate, and ship mulK-stage agenKc pipelines that orchestrate Anthropic Claude, OpenAI,
Gemini and local models in parallel, with dynamic tool-use chaining and typed intermediate outputs. The system evolves
weekly: new document types, new legal domains, new extracKon schemas — each requiring architectural adaptaKon rather
than configuraKon changes.
▸ RouKnely solve hard LLM-at-scale problems: context window saturaKon, cross-document hallucinaKon bleed,
token cost explosions on large corpora, semanKc driq between extracKon stages. Solved these not by reaching for a library
but by designing custom primiKves — Hierarchical Map-Reduce sharding, semanKc anchor engines, typed reducKon schemas
with conflict resoluKon — each validated through a custom evaluaKon harness benchmarking against naive approaches before
shipping.
▸ Maintain and extend a hybrid Computer Vision + LLM pipeline that extracts not just text but spaKal metadata:
bounding boxes, page coordinates, layout context — criKcal for court-admissible forensic document analysis. Regularly refine
the CV models, tune spaKal inference logic, and adapt the pipeline as new document formats arrive from the client.
Built ingesKon and orchestraKon endpoints in Node.js/TypeScript siwng in front of the Python-based inference
pipeline, where the hard problem was managing backpressure — accepKng bursts of document uploads without
overwhelming downstream LLM calls.
▸ Type-safe contracts (TypeScript) between the API layer and the mulK-provider LLM orchestraKon logic, so that a
schema change in one extracKon stage would fail at compile Kme rather than silently corrupKng downstream legal data.
▸ Run the full DevOps side: Docker mulK-stage builds with environment isolaKon per processing layer, CI/CD
configuraKon, infrastructure monitoring, cost tracking across LLM providers. Treat this as a real product — SLAaware, incident-
responsive, conKnuously opKmized for both performance and API spend.

AI Systems Architect | Flexito (Conversa5onal AI PlaNorms & B2B Solu5ons)
Part-5me / Remote • June 2025 – February 2026

▸ Architected a producKon-grade conversaKonal AI plalorm from the ground up — a 30+ flow orchestraKon system
covering the full B2B and B2C client lifecycle across WhatsApp, Facebook, Instagram, and Webchat. Owned the full technical
decision cycle: subscriber idenKty model, mulK-channel rouKng architecture, session state machine, AI task orchestraKon
layer, and a lightweight structured data layer serving as the system's CRM.

▸ Designed and implemented an 8-task AI orchestraKon system with typed prompt contracts, structured JSON output
schemas, and confidence-threshold-based downstream rouKng — spanning intent classificaKon, BANT signal extracKon,
senKment scoring, session summarizaKon, document validaKon, dynamic FAQ retrieval, handoff brief generaKon, and
mulKlingual input normalizaKon. Each task is independently versioned and callable from any flow layer.
▸ Engineered a BANT-based lead scoring engine with weighted aggregaKon across 4 dimensions (budget, authority,
need, Kmeline), dynamic temperature classificaKon (hot/warm/cold), and automated mulKchannel escalaKon triggers.
Integrated via structured API operaKons (read, upsert, condiKonal update) — eliminaKng webhook middleware enKrely — and
extended to external CRMs including Zoho, HubSpot, and Microsoq Dynamics depending on client stack.
▸ Designed and shipped the Smart Input System (SIS) — a plalorm-level 5-Ker input degradaKon paxern that
intercepts every no-match and no-input event across all quesKon nodes. Tier 0 performs silent AI-based response mapping;
escalaKng Kers execute confidence-gated confirmaKon, AI-driven quesKon rephrase, binary simplificaKon, and context-rich
human escalaKon. Built as a fully reusable Web Sub Flow paxern requiring zero per-flow customizaKon — eliminaKng the
single largest source of conversaKon abandonment in structured bot flows.
▸ Implemented a stateful session resumpKon system using checkpoint fields and tag-based flow detecKon —
enabling seamless mid-session recovery across file upload sequences, mulK-step document collecKon, and slotbased
appointment scheduling. Includes message-type classificaKon via custom JavaScript URL paxern matching against AWS S3
media paths (voice/image/document/video rouKng) where plalorm-naKve media fields are unavailable.
▸ Built the core orchestraKon and rouKng layer in Node.js/TypeScript, handling real-Kme message dispatch across
mulKple messaging channels with typed contracts between flow layers.
▸ Used TypeScript's type system to enforce strict input/output schemas across the 8-task AI orchestraKon pipeline,
catching contract mismatches between flow versions before runKme.
▸ Wrote Node.js middleware/services responsible for session state transiKons — the hard part being concurrency:
mulKple channels wriKng to the same subscriber session without race condiKons or lost state.
▸ Implemented async, event-driven integraKons (webhooks, API polling, CRM sync) in Node.js, where the difficulty
was ordering and retry logic under unreliable third-party APIs (WhatsApp/CRM rate limits, Kmeouts).
▸ Delivered across mulKple client verKcals including financial services, real estate, legal, healthcare, and ecommerce
— adapKng the same core architecture to radically different qualificaKon models, compliance requirements, and CRM
environments. Maintains a reusable universal template system that reduces per-client delivery Kme while preserving full
configurability at the Config layer.

Independent AI Systems Consultant / Engineer | Global clients
Freelance / Remote • Jan 2024 – Present

▸ Delivered end-to-end AI systems across Medical, Legal, and EdTech verKcals — each engagement a fresh
architecture problem spanning FastAPI backends, serverless deployments, REST integraKons, and direct SDK integraKons
tailored to the client's stack.
▸ Built mulK-step LLM reasoning systems where single-shot calls weren't sufficient: decomposed complex inference
tasks into typed reasoning stages with explicit intermediate state, validated outputs at each step, and surfaced confidence
signals to downstream logic — applied across legal clause extracKon, medical record summarizaKon, and educaKonal content
classificaKon.

▸ Constructed RAG pipelines from scratch over confidenKal, private knowledge bases across mulKple vector
backends (FAISS, ChromaDB, Pinecone) and embedding strategies — validated through a repeatable evaluaKon harness
benchmarked against keyword-search baselines and tuned per-domain unKl retrieval quality met client acceptance criteria.
▸ Regularly benchmarked and selected models per use case — latency, context window, cost-per-call, and output
quality — across Anthropic, OpenAI, Gemini, Mistral, and self-hosted open-source models, rouKng to local models where cloud
calls weren't jusKfied; consistently shipped with zero post-handoff support calls, backed by architecture docs, integraKon
guides, and runbooks wrixen for non-technical operators.

KEY ACHIEVEMENTS

▸ Built and ran a self-hosted, autonomous enterprise infrastructure (KHASH) at real scale — 187 connectors across 12
categories, 32 producKon services, and a 91-VM / 12-host self-hosted GPU fleet — proving AI systems can be delivered
without dependency on external API vendors.

▸ Reduced manual processing Kme by 60–75% on average across automaKon engagements — in the most intensive case,
compressed 40+ minutes of human legal review per document batch down to under 14 minutes of fully automated AI
processing.

▸ Designed agenKc systems that scale linearly with document volume at sub-linear cost growth — processing 100+
documents (~350k tokens) per run without hiwng context limits, via custom Map-Reduce sharding and parallel sub-
agent execuKon.

EDUCATION Bachelor's in Computer SOFT SKILLS
Science Zhytomyr Polytechnic State
University ▸ Full delivery ownership & accountability
LANGUAGES
▸ CriKcal evaluaKon of AI-generated code
Ukrainian — NaKve
English — B2+ ▸ AnalyKcal & systems thinking
▸ Autonomous 0-to-1 execuKon — no
micromanagement required