Andriy

ANDRIY CHOS
AI Systems Architect | Machine Learning Engineer
[відкрити контакти](див. вище в блоці «контактна інформація») | [відкрити контакти](див. вище в блоці «контактна інформація») | github.com/40chos/oma-portfolio | Telegram: @jizelfi |
Zhytomyr, Ukraine (Remote)

PROFESSIONAL SUMMARY
AI Systems Architect and Machine Learning Engineer with 3+ years shipping production AI infrastructure end-to-end — from self-
hosted GPU clusters and deterministic orchestration engines to the agentic systems running on top of them. Built KHASH, a
deterministic-first automation architecture now running 32 production services and 187 integrations for a 2,700-project Dutch
enterprise, and OMA, an autonomous ERP development agent built on it. Delivered 30+ production AI systems on Anthropic,
OpenAI, Gemini, and self-hosted open-source models across legal, medical, and conversational-AI domains. Deterministic-first by
design: AI is invoked only where tested, human-approved logic can't already decide the outcome. Deep practical experience with
RAG, MCP, multi-agent orchestration, and hybrid Computer Vision + LLM pipelines.

TECHNICAL SKILLS
Languages: Python, TypeScript, SQL, JavaScript, C++, C#, Go, Rust
LLM APIs & Models: Anthropic Claude, OpenAI, Gemini, Mistral, Qwen, self-hosted / local LLMs
Agentic AI: MCP (Model Context Protocol), multi-agent orchestration, tool use / function calling, autonomous pipelines
RAG & Vector / Storage: ChromaDB, FAISS, Pinecone, Weaviate, Qdrant, Neo4j, OpenAI Embeddings, Sentence Transformers,
PostgreSQL (pgvector), Redis, RabbitMQ
Frameworks: PyTorch, FastAPI, Next.js, Node.js, Pydantic, LangGraph
DevOps & Infrastructure: Docker, Kubernetes, Proxmox, Linux (systemd), AWS, GitHub Actions, CI/CD, Keycloak, Teleport, vLLM
AI Coding Tools: Claude Code, Cursor, Windsurf, GitHub Copilot, Codeium

WORK EXPERIENCE

AI Systems Architect — BadkamerVakman B.V. Mar 2026 – Sep 2026
Dutch Construction & Renovation Enterprise (2,700+ projects) · Self-employed / Remote Contract, Full-time
• Engaged to design a durable, company-agnostic system (KHASH) rather than a single-client chatbot — built to outlast a typical 5-
year vendor-locked build and be re-instantiated for any mid-sized organization.
• Own the full self-hosted infrastructure underneath it: a multi-hypervisor 12-host / 91-VM Proxmox cluster running internal PKI
(step-ca), Vault + Vaultwarden secrets, Keycloak OIDC/SSO, Teleport audited SSH, Postgres (pgvector), Redis, RabbitMQ,
Qdrant/Neo4j for vector + graph retrieval, and Grafana/Loki observability.
• Added Kubernetes-orchestrated workloads on top for services needing container-level scaling, and run a self-hosted GPU model
fleet (coding, general-purpose, embedding, reranking, vision, speech via vLLM) so the platform runs fully offline/sovereign instead
of depending on an external AI vendor.
• Designed for reliability first: generated multiple independent candidate architectures, stress-tested each against real organizational
scenarios, resolved disagreements with targeted tests rather than opinion, and converged on one fault-tolerant design —
cryptographically frozen before development began.
• Formalized the converged design as a 13-part design-decision methodology covering system context, trust boundaries,
identity/PKI/secrets, AI/inference, sovereignty & offline operation, audit & evidence, and deployment & emergency-recovery
authority.
• Defined a formal integration taxonomy (8 roles × 7 kinds — Tool Adapter, Knowledge Connector, Infrastructure Driver, Secret
Provider, and more) applied across 32 production services and 187 cataloged connectors in 12 categories.
• Built ~2,500 atomic automation primitives from the client's historical customer emails, reviewed and approved once rather than
regenerated by AI each time, combined into self-verifying composites (e.g. a backup that checksums its own output) and
deterministic routing engines that automate client triage without hallucinations.

• Proved the foundation supports genuinely new AI systems, not just automating what exists — shipped OMA (Odoo Manager Agent),
an autonomous agent turning plain-English requests into real, installed Odoo ERP module code (models, fields, views, security
rules).
• OMA's stack: custom hand-built orchestration (no LangChain/LangGraph) coordinating Manager, Build, Code-Review, and
Testing/QA components; self-hosted Qwen models (27B/30B/9B) served via vLLM, with Claude Sonnet as a cloud escalation tier;
FastAPI + Pydantic + PostgreSQL, grounded by a Neo4j knowledge graph of the live Odoo schema.
• Built the operator-facing dashboard in Next.js/TypeScript surfacing real-time system state (91 VMs, GPU fleet health) without
overwhelming the operator, and lightweight typed adapter services across the 187-connector integration layer.

AI Systems Architect — Flexito (Conversational AI Platforms & B2B Solutions) Jun 2025 – Feb 2026
Part-time / Remote, concurrent with the LegalTech contract below
• Architected a production-grade conversational AI platform from the ground up — a 30+ flow orchestration system covering the full
B2B/B2C client lifecycle across WhatsApp, Facebook, Instagram, and Webchat.
• Owned the full technical decision cycle: subscriber identity model, multi-channel routing, session state machine, AI task
orchestration layer, and a lightweight structured data layer serving as the system's CRM.
• Designed an 8-task AI orchestration system with typed prompt contracts, structured JSON output schemas, and confidence-
threshold-based downstream routing — intent classification, BANT signal extraction, sentiment scoring, session summarization,
document validation, dynamic FAQ retrieval, handoff brief generation, and multilingual input normalization. Each task is
independently versioned and callable from any flow layer.
• Engineered a BANT-based lead scoring engine with weighted aggregation across 4 dimensions (budget, authority, need, timeline),
dynamic hot/warm/cold classification, and automated multichannel escalation triggers — integrated via structured API operations
that eliminated webhook middleware entirely, extended to Zoho, HubSpot, and Microsoft Dynamics depending on client stack.
• Designed and shipped the Smart Input System (SIS) — a reusable 5-tier input-degradation pattern (silent AI response mapping →
confidence-gated confirmation → AI rephrasing → human escalation) requiring zero per-flow customization, eliminating the largest
source of conversation abandonment in structured bot flows.
• Implemented stateful session resumption (checkpoint fields, tag-based flow detection) for seamless mid-session recovery across file
uploads, multi-step document collection, and appointment scheduling.
• Built the core orchestration and routing layer in Node.js/TypeScript, handling real-time message dispatch and session-state
transitions across multiple channels — the hard part being concurrency, since multiple channels write to the same subscriber
session without race conditions or lost state.
• Implemented async, event-driven integrations (webhooks, API polling, CRM sync), where the difficulty was ordering and retry logic
under unreliable third-party APIs (WhatsApp/CRM rate limits, timeouts).
• Delivered across financial services, real estate, legal, healthcare, and e-commerce verticals — adapting the same core architecture
to radically different qualification models, compliance requirements, and CRM environments via a reusable universal template
system rather than rebuilding each client integration from scratch.

Founder, then Lead AI Engineer (Contract) — LegalTech Internal Tooling (U.S. Legal Firm) Apr 2025 – Apr 2026
Self-employed / Remote · Full-time through Feb 2026, part-time Mar–Apr 2026 during the BadkamerVakman handoff
• Brought the original product idea and an initial build to a practicing U.S. legal firm, who funded and staffed its legal/compliance
side; later divested founder stake and continued on a paid contract as Lead AI Engineer.
• Owned the full engineering lifecycle of this firm's internal AI tooling — from architecture decisions to production deployment —
across a year of active build-out followed by continuous iteration: new document types, legal domains, and extraction schemas each
required architectural adaptation, not just config changes.
• Shipped multi-stage agentic pipelines orchestrating Claude, OpenAI, Gemini, and local models in parallel, with dynamic tool-use
chaining and typed intermediate outputs.
• Routinely solved hard LLM-at-scale problems — context window saturation, cross-document hallucination bleed, token cost
explosions, semantic drift between extraction stages — with custom primitives: hierarchical Map-Reduce sharding, semantic
anchor engines, and typed reduction schemas with conflict resolution, each validated through a custom evaluation harness before
shipping.
• Built and maintained a hybrid Computer Vision + LLM pipeline that extracts not just text but spatial metadata — bounding boxes,
page coordinates, layout context — critical for court-admissible forensic document analysis; refined the CV models and spatial
inference logic as new document formats arrived.
• Built the Node.js/TypeScript ingestion and orchestration layer sitting in front of the Python inference pipeline, where the hard
problem was backpressure — accepting bursts of document uploads without overwhelming downstream LLM calls — with type-safe
contracts ensuring a schema change in one extraction stage failed at compile time instead of silently corrupting downstream legal
data.

• Ran the full DevOps side: Docker multi-stage builds with environment isolation per processing layer, CI/CD, infrastructure
monitoring, and cost tracking across LLM providers — treated as a real product: SLA-aware, incident-responsive, continuously
optimized for both performance and API spend.

Independent AI Engineer & Developer — Parallel Freelance Practice — Global Clients October 2023 – Present
Ongoing, part-time micro-engagements running continuously alongside the contracts above — not a dedicated full-time role · Freelance
/ Remote
• Delivered AI systems and infrastructure work ranging from short scoped tasks to end-to-end ownership, across Medical, Legal, and
EdTech verticals — FastAPI backends, serverless deployments, REST integrations, and direct SDK integrations tailored to each
client's stack.
• Built multi-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 extraction, medical record summarization, and educational content classification.
• Constructed RAG pipelines over confidential, private knowledge bases across multiple vector backends (FAISS, ChromaDB, Pinecone)
and embedding strategies — evaluated against keyword-search baselines and tuned per-domain until retrieval quality met client
acceptance criteria.
• Benchmarked and selected models per use case (latency, context window, cost-per-call, output quality) across Anthropic, OpenAI,
Gemini, Mistral, and self-hosted open-source models, routing to local models where cloud calls weren't justified; delivered with
architecture docs, integration guides, and runbooks written for non-technical operators.

KEY ACHIEVEMENTS
• Built and ran a self-hosted, vendor-independent enterprise AI infrastructure (KHASH) at real scale: 187 connectors across 12
categories, 32 production services, and a 91-VM / 12-host self-hosted GPU fleet.
• Reduced manual processing time by 60-75% on average across client automation engagements at BadkamerVakman and in
freelance work — the most intensive case compressed 40+ minutes of human legal review per document batch to under 14
minutes.
• Designed agentic systems that scale sub-linearly in cost with volume — processing 100+ documents (~350k tokens) per run without
hitting context limits via custom Map-Reduce sharding and parallel sub-agent execution.

EDUCATION & ADDITIONAL INFORMATION
Bachelor's in Computer Science — Zhytomyr Polytechnic State University
Languages: Ukrainian (Native), English (B2+)
Soft Skills: Full delivery ownership & accountability · Critical evaluation of AI-generated code · Analytical & systems thinking ·
Autonomous 0-to-1 execution