Independent AI engineering practice of Mikołaj Jesion

I take AI from pilot to production.

LLM systems, document AI and predictive models, built on AWS, Azure or your own servers. You work directly with the engineer who designs and builds the system. No account managers, no handoffs to juniors.

Tell me about your project See the work

~2Mdocuments processed

~50%faster software delivery

+20%sales from recommendations

Worked with

Silent Eight
BEST GK
Euvic Media
IBiMM
DataCourage
PLASMET
MindSyncown product

01 / Work

Every point is a record. Every cluster is a system I shipped.

Emails, documents, transactions, app sessions. As you scroll, the query moves through the space and connects to its nearest neighbours. There are eight clusters, and each one is a real project with a number from production.

email
document
transaction or event
query

LLMs and document AI
AI-assisted delivery
Predictive models
AI products

Cluster 01 / 08LLMs and document AI

queryEmails and attachments relevant to the dispute, in any file format

eDiscovery over about 500 GB of email, in production

I built an eDiscovery system on LLMs and AI search for a law firm, as part of my work with BEST GK. It finds what matters in about 500 GB of email and attachments in mixed formats. The firm uses it in production to prepare for a major court case.

Client: A law firm, via BEST GK
Status: In production
Built with: LLMs, AI search, PostgreSQL, Azure Blob Storage
Service: LLMs and document AI

~500 GB

Legal eDiscovery. Email and attachments, searchable in production

Cluster 02 / 08LLMs and document AI

queryTransactions an analyst should review, without the data leaving the bank

Anti-money laundering on local LLMs

At Silent Eight, I co-build an AML system that runs entirely on local models inside the bank. No data leaves its infrastructure, so nothing goes to an outside API.

Client: Silent Eight, ongoing
Role: Co-building the AML system
Data: Stays inside the bank
Service: LLMs and document AI

Local LLMs

AML on local LLMs. Runs inside the bank, no data leaves

Cluster 03 / 08LLMs and document AI

queryThe facts, the category and a short summary of every incoming document

Extraction and summaries for about 2 million documents

For BEST GK I built architectures on AWS and Azure for data extraction and summaries. They have processed about 2 million documents, and they still run as part of the company's daily work.

Client: BEST GK, ongoing
Scope: Extraction, classification, summaries
Built with: LLMs on AWS and Azure
Service: LLMs and document AI

~2M

Document AI at scale. Documents processed on AWS and Azure

Cluster 04 / 08AI-assisted delivery

queryShip software faster with AI agents, and still see how they are used

Software delivery about 50% faster, on Claude

At BEST GK I led the move of software delivery to Claude: an agent harness, training for the teams, and oversight with metadata from every conversation stored in ClickHouse. Real end-to-end development tasks now take about 50% less time, measured with Claude Code versus without.

Client: BEST GK, ongoing
My role: Led the change, end to end
Built with: Claude, agent harness, ClickHouse
Service: AI-assisted software delivery

~50%

AI-assisted delivery. Less time on real end-to-end tasks, with Claude Code vs. without

Cluster 05 / 08Predictive models

queryProducts this shopper is likely to want next

Recommendations that added 20% to sales

At Euvic Media I was the only data scientist on the engagement and built the recommender end to end, for more than 120 online stores, using collaborative and content-based filtering. The result: +20% sales across all 120+ stores where the recommender was deployed.

Client: Euvic Media, completed
Reach: More than 120 online stores
Built with: FastAPI, MongoDB, ClickHouse, Redshift
Service: Predictive models and recommendations

+20%

Recommendations. Sales across all 120+ stores with the recommender

Cluster 06 / 08Predictive models

queryWhat is this ad impression worth to us, right now?

A real-time bidding system that raised margins

Also at Euvic Media, I built a real-time bidding system end to end, with three predictive models behind every bid. It raised margins by up to several tens of percent across many campaigns.

Client: Euvic Media, completed
Models: Three, one decision per bid
Result: Higher margins on many campaigns
Service: Predictive models and recommendations

Tens of %

Real-time bidding. Margins up by up to several tens of percent

Cluster 07 / 08Predictive models

queryWhich children show early signs of a posture defect?

Nine medical classifiers, all above 0.85 AUC

At IBiMM (Institute of Biotechnology and Molecular Medicine) I worked as a senior data scientist on PostureScan, early detection of posture defects in children. I built 9 classifiers for orthopedic conditions, including scoliosis. All of them reached AUC above 0.85.

Where: IBiMM, PostureScan project
Domain: Pediatric orthopedics
Models: 9 classifiers, incl. scoliosis
Service: Predictive models

9 models

Medical classifiers. Every one with AUC above 0.85

Cluster 08 / 08AI products

queryCan one engineer ship the whole product, from model to marketing?

Two products built end to end

MindSync is my own product, where I'm CTO: an app with an AI agent that holds real-time voice sessions in 7 languages, with more than 2,000 downloads. For PLASMET I run the whole online store: design, SEO, campaigns, blog and customer email drafts, automated with Claude. Google clicks went up about 2.5x after the migration.

MindSync: Own product, CTO
PLASMET: About 7,900 products
Built with: Flutter, FastAPI, realtime voice, Claude
Service: AI products end to end

2,000+

MindSync. App downloads

~2.5x

PLASMET. Google clicks after migration

02 / Numbers

The numbers behind the clusters

Each figure comes from one project and keeps its qualifier. Ask me about any of them on a call.

01 / Document AI

~2M

documents processed

BEST GK, AWS and Azure

02 / eDiscovery

~500 GB

eDiscovery corpus in production

A law firm, via BEST GK

03 / Delivery

~50%

faster software delivery

BEST GK, Claude Code vs. without

04 / Recommendations

+20%

sales, 120+ stores

Euvic Media, recommendations in every deployed store

05 / Real-time bidding

Tens of %

higher margins, up to several tens of percent

Euvic Media, across many campaigns

06 / Medical

medical classifiers, all with AUC above 0.85

IBiMM, PostureScan

07 / MindSync

2,000+

app downloads

My own product, 7 languages

08 / PLASMET

~2.5x

Google clicks after migration

Store, SEO and content automated with Claude

09 / Experience

years of Python, 6+ in data science

Legal, compliance, finance, retail, ad tech, health

03 / Services

Four kinds of systems I build

This is the whole map: each service is a group of clusters you just flew through.

LLMs and document AI

Search, extraction, classification and summaries over documents and email. On the cloud you already use, or on local models when the data cannot leave your servers.

eDiscoveryAML~2M docs

AI-assisted software delivery

An agent harness for your developers, training for the teams, and an oversight layer that shows how the tools are used.

ClaudeClickHouse~50% faster

Predictive models and recommendations

Recommendations, bidding, risk and propensity scoring, medical classifiers. Each model is measured against one metric agreed up front.

+20% salesRTBAUC > 0.85

AI products end to end

Architecture, backend, the mobile or web app, the AI agent inside it, and the SEO and content pipeline that brings users in.

MindSyncPLASMET

Ways to work with me

AdvisoryUse case review, architecture and vendor decisions.

Fixed-scope projectA proof of concept or a full build, with one agreed metric.

Embedded contractorB2B, inside your team and your repositories.

Fractional AI leadA fixed number of days a month as your AI lead.

For larger scopes, I bring in senior engineers I have worked with before. I stay accountable for architecture and delivery, and you keep one point of contact.

04 / About

Mikołaj Jesion

Founder of NeuraVec / AI engineer and data scientist / Poznań, Poland

I've written Python for more than 9 years and built data science systems for more than 6, in legal, compliance, finance, e-commerce, ad tech and healthcare. I trained as a systems engineer in control systems at Wrocław University of Science and Technology.

NeuraVec is my independent practice. When you hire it, you work with me: I scope the system, write the code, put it into production and hand it over to your team.

My full background is on LinkedIn.

Now
NeuraVecFounder. Independent AI engineering practice
Ongoing
Silent EightCo-building an AML system on local LLMs
2024
MindSyncCTO of my own product: AI agent, real-time voice, 7 languages
2023
IBiMMSenior Data Scientist. PostureScan, 9 medical classifiers
2022
BEST GKSenior Data Scientist. Document AI, eDiscovery, AI-assisted delivery
2022
Euvic MediaSenior Data Scientist. As the only data scientist, I built the recommender and the RTB system end to end
2020
DataCourageData Science Team Leader. Forecasting, churn models, NLP
2020
MSc, IT in BusinessData analysis and IT solutions for business problems
2016
Wrocław University of Science and TechnologyBEng in systems engineering, control systems

05 / Contact

Tell me about your project

Describe the data and what you need from it. The message comes straight to me, and I'll tell you whether it's worth a proof of concept.

Where the data lives today: on-premises, Azure, AWS or other
Whether it may leave your environment
Who would own the system once it is built

Prefer email? Write to hello@neuravec.com.

Name

Company

Your project (a few sentences is enough)

I reply within 24 hours.

Thank you. I have your message.

It went straight to my inbox. I will reply by email within 24 hours.