RKRoKenshiApplied AI EngineeringStart a conversation
Based in Singapore / Bangkok · open to remote B2B engagements

Senior Applied AI / AI Platform Engineer

AI systems that survive contact with production.

I take RAG, agent and LLM systems from an ambiguous workflow to architecture, production code, evaluation, deployment and handover.

Hire me on contract Discuss an AI build
See end-to-end reference builds
6+ yearsproduction engineering
3 buildsend-to-end reference delivery
1,610selected tests across the builds
SG / BKKUTC+7 / UTC+8 · remote B2B

01 / End-to-end reference builds

What I can deliver—quickly and end to end.

Three focused builds demonstrate the scope I can own: product framing, architecture, implementation, evaluation, deployment and operational handover.

WHY THESE BUILDS EXIST

They are not boxed products being resold. Each is a compact, full-cycle example of how I turn a real AI problem into working software in a short, focused delivery cycle—with measurable evidence and explicit limits.

01

REFERENCE BUILD · AIP PRIVATE KNOWLEDGE

Evidence-first multimodal RAG for private knowledge

Documents, calls, video and Telegram become one project-scoped knowledge layer where every answer remains traceable to an exact source passage.

0.50 → 1.00Recall@2 · five-query golden subset
6 / 6live evidence cases with required sources
−86%retrieval time · one controlled anchor case
Architecture, stack and claim boundary
SYSTEM

PostgreSQL source of truth → MinIO originals → Redis/ARQ jobs → Qdrant + BM25 + pg_trgm → fusion, MMR and reranking → grounded answer and citations.

CLAIM BOUNDARY

Controlled project benchmarks on small evaluation lanes. The latency result is one exact-anchor case, not a production SLA.

PythonFastAPIPostgreSQLQdrantRedis/ARQMinIO/S3MCP
Live deploymentOpen the live AI workspace
02

REFERENCE BUILD · AGENT FACTORY

A durable control plane for delegated coding agents

Bounded context, managed Git worktrees, ownership leases, restart recovery and independent acceptance evidence turn delegation into a reviewable engineering workflow.

9 / 9adaptive policy passes vs 7 / 9 control
−84.8%median input tokens
−41.2%p95 latency in the component A/B
Architecture, stack and claim boundary
SYSTEM

MCP Director → durable control plane → approved task DAG → provider-neutral workers → managed worktrees → deterministic checks → signed Eval Capsule.

CLAIM BOUNDARY

Seeded 18-call synthetic A/B: three cases, three repeats, two policies, one model/provider/date. It validates one context-preparation mechanism.

PythonFastAPIMCPGit worktreesEd25519OIDCPostgreSQL
Live deploymentOpen the live control plane
03

REFERENCE BUILD · ENGLISH BRO

Conversational AI with deterministic learner state

A Telegram text-and-voice coach separates evidence extraction, deterministic state updates, pedagogical policy and structured generation—then makes every turn replayable.

97.2%strict validity · separate semantic audit
71 / 71deliverable shadow-replay responses after repair
$0.00164consolidated evaluation spend per turn
Architecture, stack and claim boundary
SYSTEM

Lambda ingress → SQS FIFO → evidence-only observation → deterministic reducers → policy → typed generation → validation → outbox → Neon + S3 archive.

CLAIM BOUNDARY

Replay evidence, not a long-term learning-outcome claim. Validity, latency and cost come from related but separate evaluation stages.

PythonPydanticAWS LambdaSQS FIFONeonDeepgramOpenRouter
Live deploymentOpen the live product

02 / Engagement

One engineer. Two clear ways to work.

Choose the path that matches the ownership you need.

03 / Services

Start at the right level of certainty.

Planning ranges become a final scope only after qualification.

01

AI Architecture & Feasibility Sprint

Turn an AI idea into a decision-ready architecture, data and risk inventory, build-vs-buy view, evaluation strategy and pilot plan.

Typical: 1–2 weeksBest when the use case is real but the right system is still unclear.
02

Evidence-first RAG Pilot

Ingestion, hybrid retrieval, citations, abstention, access scope, frozen evaluation, telemetry, deployment and handover.

Typical: 4–8 weeksBest for private knowledge, document intelligence, research or support.
03

Reliable Agent Workflow

Typed tools or MCP, permission boundaries, durable state, idempotency, approvals, verification, routing, cost accounting and runbooks.

Typical: 4–10 weeksBest when an AI system must take actions—not merely answer questions.
04

AI Reliability Audit

A focused review of an existing RAG or agent prototype: failure modes, eval coverage, observability, security, latency and cost.

Typical: 3–5 daysBest when a demo works but the route to production is uncertain.

04 / How I ship

Every stage must reduce uncertainty or produce evidence.

  1. 01

    Discover

    Users, workflow, data, failure cost and the baseline that matters.

  2. 02

    Build

    A thin end-to-end slice through real interfaces before broad scope.

  3. 03

    Evaluate

    Versioned examples, regression gates and explicit release criteria.

  4. 04

    Operate

    Tracing, budgets, failure policy, deployment, rollback and runbooks.

A useful boundary

What I will not automate blindly

High-impact actions without scoped permissions, idempotency, verification and an accountable human decision path. Reliability is a product requirement, not a cleanup phase.

05 / About

Backend discipline, applied to probabilistic systems.

I’m Dmitry Popov, a senior engineer based between Singapore and Bangkok, working remotely across UTC+7 and UTC+8.

My path runs through freight logistics, enterprise asset management, fintech and insurance. That background shapes how I build AI: clear contracts, durable state, measurable quality, controlled failure and a route back from every deployment.

I’m strongest where AI meets a real backend—data ownership, async workflows, Kafka, PostgreSQL, cloud deployment, security boundaries and the operational work that makes a feature trustworthy.

Request the full résumé
2025—NOWInsurance

Lead Software Engineer · business-critical services

2024—2025Fintech

Java Software Engineer · payment platform

2022—2024Enterprise systems

Java Software Engineer · enterprise asset platform

2019—2022Transport logistics

Java Software Engineer · multimodal freight systems

PYTHON · FASTAPI · JAVA · SPRING · POSTGRESQL · KAFKA · AWS · KUBERNETES · RAG · MCP · EVALS · LLMOPS

06 / FAQ

Before we talk.

Direct answers to the questions that usually matter first.

01Can I hire you into an existing product team?

Yes. I work as a senior remote B2B engineer and can own an AI feature or platform area from design through production operation and knowledge transfer.

02What systems do you build?

Private RAG and document intelligence, tool-using agents, MCP integrations, conversational products, evaluation systems and the backend platform services behind them.

03How do you measure AI quality?

With versioned golden datasets, deterministic and live evaluation lanes, regression gates, failure analysis, latency, reliability and provider-cost metrics. The method depends on the risk of the workflow.

04Can you integrate AI into an existing Java platform?

Yes. My background combines Java, Spring, Kafka and Kubernetes systems with Python, FastAPI, PostgreSQL, vector retrieval and cloud deployment.

05Do you guarantee LLM accuracy?

No responsible engineer can promise universal accuracy from a probabilistic model. I define evaluation slices, abstention and approval policies, measurable gates and safe behavior outside the tested boundary.

06Can you quote a project immediately?

A useful estimate needs a bounded workflow, available data, integrations, risks and acceptance criteria. A short architecture sprint is the fastest route when those are not yet clear.

HAVE A TEAM OR A SYSTEM THAT NEEDS TO SHIP?

Start with the outcome, not the model.

Send the role or the problem. I’ll reply with fit, key risks and a concrete next step.

JOIN YOUR TEAMHire me on contract

Role · team · 90-day outcome · timezone · start date

BUILD WITH YOUDiscuss an AI system

Workflow · data · integrations · outcome · deadline

rouronikenshi@gmail.comTelegram ↗GitHub ↗