Open to Forward Deployed Engineer & Data/AI platform roles

Sviat Nahirnyi — Forward Deployed Engineer, Data & AI Platforms. From the warehouse floor to production AI.

I embed with the people who run operations — in warehouses, factories and client offices — map how the work really happens, and ship the Data & AI platforms they end up relying on: Lakehouses, real-time streaming, RAG and LLM evaluation.

Get in touch Download CV PDF Ask my CV

Top-20European logistics co. — the largest deal I led at the company I co-founded
9client engagements delivered by the practice I built
1 monthPoC that won an aerospace data-platform tender
6 yrsin data engineering · M.Sc. in AI

Selected clients & engagements

Dow Jones
Top-20 European logistics co.
Cisco (observability PoC)
Fortune 500
Aerospace (tender PoC)

01impact.metrics

Proof, not adjectives.

Six numbers from six years of shipping data platforms into real operations. Every one of them is on my CV and came from a production system, not a demo.

0→17
Graphit
engineers in the Data & AI practice
Built from zero across 9 client engagements — real-time platforms, Lakehouse migrations, LLM evaluation systems.
−30%
Graphit
flagged chatbot responses
Architected an LLM-as-a-Judge evaluation platform for production chatbot responses.
39→1
Sigma Software
sources into one Lakehouse
Source-of-truth platform for an international logistics company, serving 5 analytics & ML teams.
+6%
Graphit
operational performance
Embedded on-site in a top-20 European logistics company’s warehouses, then shipped the analytics platform.
−37%
Sigma Software
end-to-end pipeline runtime
Built incremental PySpark pipelines on Databricks.
50K+/s
SoftServe
events per second, real time
Spark Structured Streaming for telecom network-map visualisation and latency detection.

02deployment.log

Six years, two tracks, one job: make data useful.

I co-founded Graphit in 2023 and ran its Data & AI practice while staying hands-on in client delivery — so from 2023 the tracks run in parallel.

From March 2023 the tracks run in parallel: co-founding Graphit while staying hands-on in client delivery. Select a bar to jump to the full role.

Founder track · Mar 2023 – Aug 2026

Co-Founder & Lead, Data & AI Platforms · Graphit (opens in a new tab) (now Datasoft Group)

London, UK

Co-founded a London data & AI consultancy and built its Data & AI practice from zero — selling, designing and shipping platforms on-site with clients.

Closed the company’s largest enterprise deal — a top-20 European logistics company — by leading the pitch and end-to-end solution design.
Embedded on-site at the client’s warehouses to map operational workflows, then delivered analytical data platform solutions that improved operational performance by 6%.
Worked forward-deployed at Dow Jones’ Barcelona office on vector database scalability for production retrieval workloads.
Grew the Data & AI practice from zero to 17 engineers across 9 client engagements — real-time platforms, Lakehouse migrations, LLM evaluation systems.
Cut flagged chatbot responses by 30% by architecting an LLM-as-a-Judge evaluation platform.

LLM-as-a-Judge
Vector retrieval
Lakehouse migrations
Real-time platforms
Presales
Solution design

Jul 2023 – Jan 2025Münster, Germany
Senior Data Engineer · Sigma Software Group
- Delivered the source-of-truth Lakehouse for an international logistics company — 39 operational sources consolidated into one platform serving 5 analytics and ML teams — by embedding on-site at the client’s factory to design ETL flows with operations staff.
- Designed analytical dashboards for factory and logistics stakeholders on-site, mapping shop-floor workflows directly into data models and reporting.
- Reduced end-to-end pipeline runtime by 37% by building incremental PySpark pipelines on Databricks.
- Databricks
- PySpark
- Lakehouse
- ETL
- Dashboards
May 2023 – Aug 2023Hybrid, US
Presales Software Engineer · GreenM
- Won a data-platform tender for an aerospace company via a one-month PoC delivered directly with the client’s stakeholders, validating ingestion, storage, and query latency on terabytes/day.
- Enabled unified alerting across 7 networking-metric sources by building the Kubernetes-native ingestion layer for a Cisco observability PoC.
- Kubernetes
- PoC delivery
- Observability
- TB/day ingestion

Earlier roles · 2020–23Data Engineer · Netminds · Sep 2022 – May 2023Big Data Developer · SoftServe · Sep 2021 – Sep 2022Data Engineer · N-iX · Oct 2020 – Sep 2021

Sep 2022 – May 2023Remote
Data Engineer · Netminds
- Halved new-pipeline delivery time (4 days → 2) by rebuilding the platform around reusable ingestion templates.
- Lifted loyalty engagement 10% with Spark / Scala pipelines feeding an AI-driven offers engine.
- Shortened release cycles from 3 days to 1 by establishing Azure DevOps CI/CD for Databricks.
- Spark
- Scala
- Databricks
- Azure DevOps
Sep 2021 – Sep 2022Remote
Big Data Developer · SoftServe
- Built a Spark Structured Streaming system handling 50K+ events/second for real-time telecom network-map visualisation and latency detection.
- Translated ambiguous stakeholder requirements into service boundaries, SLAs, and data contracts; built demos used in design reviews that helped upsell the account.
- Spark Structured Streaming
- Real-time
- Data contracts
- SLAs
Oct 2020 – Sep 2021Remote
Data Engineer · N-iX
- Secured sensitive datasets for a Fortune 500 company with a permissioned AWS data warehouse — Airflow, RBAC, Terraform CI/CD — plus a data-quality framework of 100+ automated checks.
- AWS
- Airflow
- Terraform
- RBAC
- Data quality

Education M.Sc. & B.Sc. in Artificial Intelligence — Lviv Polytechnic National University

03case.studies

Two of many deployments, up close.

Six years of client delivery, nine engagements at Graphit alone. Here are two of my most recent projects, end to end: an AI system and a data platform — each with its live pipeline.

Case 01 · AI · Graphit

Cutting flagged chatbot responses by 30%.

I architected an LLM-as-a-Judge evaluation platform that checks a chatbot’s answers twice: live, before a user sees them, and offline, before every release. Flagged responses fell by 30%.

Read the full case study

case 01 · chatbot_answers → judged_answers

demopausedstatic frame · reduced motionsimulated≈ 1,840 answers/minflag rate ≈ 10%−30% flagged

the bot’s sources
- docs
- tickets
- chats
- users
chunk · vectorise
vector index · top-k
LLM · answer
live gate · every answer
- users

document chunk
query
retrieved context
answer
flagged by judge

Hover or focus a stage for detailsTap a stage for details

A live demo of the pipeline, rewired for this case. The chatbot side is simplified and the judge’s criteria are illustrative; the judge, the eval set and the release loop are what I built. Rates are simulated — the −30% is the real result.

04how.i.work

Forward deployed, end to end.

The job isn’t just writing pipelines. It’s earning trust on-site, finding the real problem, shipping something that runs in production — and proving it worked.

with DAG("forward_deployed") as dag:
    embed >> map >> ship >> prove >> scale

01
Embed
Sit with the operators where the work happens — not in a requirements meeting.
- Warehouses, a factory floor and Dow Jones’ Barcelona office
02
Map
Turn tribal knowledge into data models, contracts and SLAs that engineers can build against.
- Shop-floor workflows → data models, contracts & SLAs
03
Ship
Build the platform in production — streaming, Lakehouse, retrieval — not in slides.
- Lakehouse, streaming and retrieval in production
04
Prove
Measure it: evaluation gates, data-quality checks and the business KPI that moved.
- LLM-as-a-Judge evals · 100+ data-quality checks
05
Scale
Turn one success into the next deal — and the team to deliver it.
- Largest enterprise deal · a tender won with a PoC

Sviat Nahirnyi speaking with a microphone on a panel at the Data & AI Summit — fig. field_note
Same job on a panel as on a warehouse floor: make Data & AI *useful* to the people who have to live with it.
Panelist · Data & AI Summit

05rag.playground

Ask my CV. Watch the retrieval.

It runs 100% in your browser — hybrid retrieval over 44 chunks of my CV, then an eval gate decides whether the answer is allowed out — a small, rule-based take on the LLM-as-a-Judge idea behind my −30% flagged-responses result.

ask_cv · console

44 chunks · 384-d int8 · 49 KB

Try

~29 MB · opt-in

Off · BM25 keyword retrieval, instant. Suggested questions already use precomputed embeddings. On downloads ~23 MB model + ~6 MB runtime, once.

Answer · extractive

idle

Pick a question above or type your own. The answer is stitched together from CV sentences only — then the judge decides whether it may be shown.

fast · BM25 · k=4

Embedding map · 2D PCA

44 chunks

AI / LLM12
Data & Streaming9
Forward Deployed11
Cloud & DevOps4
Leadership2
Profile6

MiniLM vectors projected to 2D; the query lands near what it retrieves. A sketch — 2D keeps ~20% of the variance, retrieval uses all 384 dimensions.

Eval gate · judge

idle

5 checks · awaiting a question

Retrieval relevance—awaiting a question
Groundedness—awaiting a question
Citation coverage—awaiting a question
Scope—awaiting a question
Privacy policy—awaiting a question

Judge strictness50 · balanced

relevance ≥ 0.35 · grounded ≥ 70% · coverage ≥ 60%

Retrieved chunks · top-4

k=4 of 44

Nothing retrieved yet.

how_it_worksWhat runs where — and what’s real

query
tokenise
BM25 + MiniLM 384-d cosine
top-k (k=4)
extractive answer
judge
verdict

Index offline · Node: 44 chunks generated from the same data as this page → all-MiniLM-L6-v2 (quantised, mean-pooled, normalised) → 384-d vectors stored as int8, plus a 2D PCA basis. Ships as a 49 KB JSON that loads only when you scroll here.
Fast mode your browser: BM25 (k1 1.2, b 0.75) with stopwords, light stemming and a small domain synonym map — llm → judge, rag, evaluation; cloud → aws, azure, terraform. Instant; no model.
Suggested questions your browser: Their embeddings were computed at build time, so they get true hybrid retrieval (normalised BM25 + cosine) without downloading anything.
Semantic mode your browser · Web Worker: Only when you switch it on: transformers.js and the same MiniLM model (~23 MB + ~6 MB runtime, cached afterwards). Your question is embedded locally and compared by brute-force cosine with all 44 chunks — right at this size; at production scale that becomes an ANN index such as HNSW.
Answer extractive: The 2–3 best sentences from the top-4 chunks, ranked by query-term overlap × chunk score, each with a citation. No generated text, so nothing to hallucinate.
Judge rule-based: A deterministic stand-in for an LLM-as-a-Judge with the same gate: relevance, groundedness, citation coverage, scope and a privacy policy. Extractive answers are grounded by construction — the check is there so a generative model could never slip an unsupported claim past it.
Map decorative: 2D PCA keeps ~20% of the variance, so treat it as a sketch. Typed questions without an embedding are placed at the score-weighted centroid of their top-4.
Privacy by design: Nothing you type leaves this tab — there is no server and no LLM API behind this console.

06side.projects

Side projects, shipped.

What I build outside client work: products where AI does one useful job — analysing conversations, parsing messy documents, or extracting structure. One of them grew into a Graphit platform.

AI analytics · Side project → Graphit platform
ConvoInsights
See where your chatbot fails — in every conversation.
Analyses every AI-chatbot conversation to surface drop-offs, failure flows and frustrated users — topics, sentiment and recommendations that help development and CX teams cut escalations.
Started as a side project; it later became one of Graphit’s key platforms for its customers.
- Conversation analytics
- LLM classification
- Sentiment
- Dashboards
iOS app · AI
Fortium (opens in a new tab)
The program you already follow. Now it runs itself.
A strength-training log that turns a coach’s PDF, a spreadsheet or a photo of a notebook into a working program — with progressive-overload targets on every set, PR detection and an AI coach that proposes changes for your approval.
- Document → structured data
- LLM parsing
- iOS
App · AI
Actium (opens in a new tab)
Turn self-help books into daily quests.
AI that extracts the lessons from a book and turns them into personalised daily actions — with streaks, points and progress across Body, Mind, Soul and Spirit.
- LLM extraction
- Personalisation
- Gamification
Open source · MCP
Hilka (opens in a new tab)
Log how you think, not just what you decided.
A decision-tree journal that keeps the branches you rejected. Ships an MCP server and a Claude Code plugin so AI assistants can write decision trees straight into your account.
- React 19
- Supabase / Postgres
- Cloudflare Workers
- MCP
source on GitHub (opens in a new tab)

07stack.lock

Tools I reach for.

No skill bars — the tools I work with, most of them on client projects.

01 The part that doesn’t fit on a tech list.

Forward Deployed

On-site solution design
Enterprise PoCs
Requirements → data contracts
Presales

02 Retrieval, evaluation and agentic systems.

AI / LLM

HuggingFace
LangChain
LLM-as-a-Judge evaluation
Vector retrieval
RAG
MLflow
Agentic systems

03 The platforms everything else stands on.

Data & Streaming

SQL
Spark
Streaming
Databricks
Airflow
Kafka
Python
Scala
Flink
Snowflake

04 Shipped, versioned, repeatable.

Cloud & DevOps

AWS (SageMaker, Bedrock, Glue, EKS, S3)
Azure (ADF, Synapse)
MLOps
Kubernetes
Terraform

08deploy

Got messy operational data and an AI ambition?

I’m looking for my next Forward Deployed Engineer or Data & AI platform role — somewhere I can sit with customers, find the real problem and ship the thing that fixes it. London-based, happy to travel on-site.

Email me

Start a conversation LinkedIn Download CV PDF · 104 KB

Available now · London, UK
UK right to work — no sponsorship needed
Hybrid · on-site · remote · relocation

Sviat Nahirnyi — Forward Deployed Engineer, Data & AI Platforms. From the warehouse floor to production AI.

Proof, not adjectives.

Six years, two tracks, one job: make data useful.

Senior Data Engineer · Sigma Software Group

Presales Software Engineer · GreenM

Data Engineer · Netminds

Big Data Developer · SoftServe

Data Engineer · N-iX

Two of many deployments, up close.

Cutting flagged chatbot responses by 30%.

39 sources, one source of truth.

Forward deployed, end to end.

task_id=Embed

task_id=Map

task_id=Ship

task_id=Prove

task_id=Scale

Ask my CV. Watch the retrieval.

Answer · extractive

Embedding map · 2D PCA

Eval gate · judge

Retrieved chunks · top-4

Side projects, shipped.

ConvoInsights

Fortium (opens in a new tab)↗

Actium (opens in a new tab)↗

Hilka (opens in a new tab)↗

Tools I reach for.

Forward Deployed

AI / LLM

Data & Streaming

Cloud & DevOps

Got messy operational data and an AI ambition?

Embed

Map

Ship

Prove

Scale

Fortium (opens in a new tab)

Actium (opens in a new tab)

Hilka (opens in a new tab)