Line 19
AI engineering · Architecture · Delivery

We turn AI initiatives into production systems.

We work with product and engineering teams to define the architecture, make the hard technical decisions, build the critical pieces, and get AI into production.

AI architecture · RAG · Agents · Evaluation · Model systems · Local AI
Start a project See how we work
Production architecture01 / 03
Company knowledge

“Why did we change the authentication flow?”

Cutaway
SourcesGitHubLinearNotionConfluenceSlack
1Permission filteraccess lists applied before retrieval
2Hybrid retrievalidentifiers and meaning, then rerank
3Grounded answerno source, no claim
Key decisionPermissions are filtered before retrieval, not after.
5 sources·citations required·every answer tracedArchitecture & decisions
Healthtech product

“What should I make of this pattern in my symptoms?”

Cutaway
InputsOn-device historyReviewed corpus
1Consent gatechecked on the server, not in the app
2Aggregate contextraw entries stay on the phone
3Validated replyreading level is a release gate
Key decisionThe model never sees raw health data.
on-device analysis·reviewed sources·nothing identifying loggedArchitecture & decisions
Infrastructure agent

compare staging and production schemas

Cutaway
EnvironmentsDevStagingProduction
1Scope checkoauth scopes, checked per tool call
2Operationscompare, migrate, export
3Audit logwho ran it, state before and after
Key decisionProduction is view-only by default.
up to 8 registries·scoped per call·every operation loggedArchitecture & decisions
United States & Europe
We’re where you are: a team on both sides of the Atlantic.
Open source
Audit how we build before you hire us.
Senior team
The software engineers you meet write the code.
Direct access
We work inside your team: your repo, your Slack, your review process.
Our team’s experience includes
What we build

Three things, built to run in production.

We design, build and ship AI systems, forward-deployed: embedded with your team, on your data and infrastructure, until it runs in production. Everything lands in your repositories, reviewed together, so nothing depends on us afterwards.

AI products

AI features built into the products people already use.

Search, copilots, voice, multimodal input, personalization. We build the application around the model, not just the endpoint.

Knowledge & RAG

Private knowledge systems connected to your own data.

Ingestion, retrieval, reranking, permissions, citations, evaluation. Long-context caching where retrieval is the wrong tool.

Agents & automation

Agents that work across engineering and business systems.

Automated development pipelines on local models, code review agents, orchestration, MCP servers, and infrastructure access with human approval gates.

The model is one layer. We build the rest.

Most AI features stall between the demo and production because the model is the easy layer. The hard part is the retrieval, tools, evaluation, API and interface around it. We own every layer, so the system reaches production instead of stopping at the prompt.

Already chosen a model vendor or platform? We build the rest around it.

  1. ModelsHosted or local, chosen per task
  2. Retrieval & toolsRAG, MCP, grounding
  3. Agents & orchestrationRouting, retries, approvals
  4. Evaluation & guardrailsTest sets, safety, cost
  5. APIs & interfaceWeb, mobile, internal tools
  6. InfrastructureDeploy, observe, scale
Systems

Three systems, and the decisions behind them.

How we design production AI: the architecture, the stages, and the trade-off we made at each one.

01

Company knowledge, with permissions.

Fewer repeated questions to senior people, and no leaked documents. Ask across GitHub, Linear, Notion, Confluence and Slack; get an answer that cites its sources and shows only what you are allowed to see.

Covers: ingestion · query understanding · retrieval · permissions · citations · evaluation · observability
Decisions
Revoked access is revoked in minutes, not at the next full sync.

A document someone lost access to yesterday must not show up today. Connectors sync incrementally and carry permission changes as first-class events.

Permissions are filtered before retrieval, not after.

Post-filtering leaks: a top-10 list with 8 items removed tells the asker something exists. We carry the access list into the index and filter there.

Hybrid search over pure vectors.

Developers search for identifiers: ticket numbers, function names, ADR codes. Embeddings are bad at exact tokens. Keyword search catches them, vectors catch the paraphrases.

Citations are mandatory, refusal is allowed.

If retrieval returns nothing relevant, the system says so. An answer without a source is worse than no answer.

Not every question needs every source.

Searching everything for everything is slow and noisy. A cheap first pass rewrites the question and picks the sources; an acronym list from your own wiki does more than a bigger model.

Evaluation is part of CI.

Questions people actually asked become the test set. Changing how documents are split, or the model, without running it is a regression waiting to happen.

Every answer is traceable and priced.

When an answer is wrong, the trace shows whether retrieval, ranking or the model failed. When the bill grows, the trace shows which questions cost the most.

* Select a stage to see the decision behind it.

Ingest pipelineStoresAnswer pathQualitySources1syncConnectors2documentsChunker3chunksEmbedderaclbm25vectorsACLKeywordsVectorsAsker4questionRewrite5querySearch6candidatesRerank7top-kContext8promptLLM9structuredGuardrail10Eval setscores everychangeTracestrace per answer
02

A healthtech product with a hard privacy boundary.

AI inside a regulated consumer health app: symptom tracking analyzed on the device, and an assistant that answers from reviewed sources. The user's raw health data never reaches the model, enforced on the server rather than promised in the app.

Covers: consent gating · safety routing · context assembly · curated corpus · structured output · in-app delivery · monitoring
Decisions
The model never sees raw health data.

Analysis of the user's own patterns runs locally and deterministically. The model receives reviewed content and, with consent, an aggregate digest. The boundary is enforced by the API, not by the app.

Consent and quota are server rules.

Client-side gating is a suggestion. The API rejects any message without persisted consent and ignores entitlements the client claims for itself.

Reviewed answers before generated ones.

Most questions repeat. A cached, reviewable answer beats a fresh generation on cost, latency and safety. The model only handles what the presets cannot.

Readability is a release gate.

The audience includes people with brain fog. Replies must hit a reading-level target and a paragraph structure. The server repairs shape, and a script scores every corpus change.

The corpus is versioned and reviewed like code.

Every passage the model can quote has an author, a reviewer and a version. A change to the corpus is a pull request, not an edit in a CMS.

Nothing identifying reaches the logs.

Monitoring counts questions, routes, latencies and refusals. It never stores what a person asked. If a regulator or a breach reads the logs, there is nothing to find.

* Select a stage to see the decision behind it.

On the deviceReviewed corpusAnswer pathOpsraw data never crossesMobile app1question + digestReviewapprovedCorpuspassagesEmbeddervectorsIndexpreset matchSafety2screenedMatch3escalateRetrieve4passagesContext5promptLLM6ValidatorcheckMonitor
03

An agent with hands on production infrastructure.

Open source · MIT

Developers operate infrastructure in plain language without anyone getting write access to production. An MCP server that lets an agent run Kafka Schema Registry across environments, with production locked to view-only.

Covers: MCP · Kafka · multi-environment · authorization · migrations · audit logging
kafka-schema-reg-mcp on GitHub
> compare staging and production schemas Found 18 subjects. 3 incompatible changes: ▲ user-profile-value ▲ payment-event-value ▲ notification-settings-value production: viewonly · no changes applied
Decisions
Fewer tools, better answers.

A slim mode cuts the tool set from more than fifty to about nine. Smaller tool lists mean fewer wrong picks and a much cheaper context on every call.

Authorization is scope-based, not prompt-based.

Read, write and admin scopes are checked per tool call. Telling the model to be careful is not a security control.

Production is view-only by default.

An agent that can read production and write to staging is useful. An agent that can write to production is an incident. The flag lives on the server, per registry, where the model cannot reach it.

Migrations back up first and verify after.

Moving schemas between registries is the operation most likely to go wrong. So it snapshots before, applies, and compares after.

Every operation is logged, before and after.

Compare, migrate, export: each call records who ran it, which registry, and the state on both sides of the change. Nothing the agent does to production infrastructure happens off the record.

* Select a stage to see the decision behind it.

MCP clientLine 19 · MITRegistries · up to eightAuditDeveloperAgent1MCPMCP server2scopesTool setpickOperations4loggedAudit log3per registrydevstagingproductionmigrate · verified
Who you work with

A senior, efficient team.

No account managers, no hand-offs. The people you talk to are the people who build it.

Eugene Zharkov
United States

Eugene Zharkov

Product & AI engineering

Puts AI in front of users. Agents, retrieval, the APIs under them and the apps on top, built as one system. Products in the hands of tens of millions of people.

AI product integration · Agents & tooling · Mobile · Web · Platform architecture
GitHub
Roman Melnyk
Europe

Roman Melnyk

Backend & AI infrastructure

Builds the machinery behind the AI. Pipelines where local models write, review and ship code on their own. MCP servers that let agents touch production without breaking it. Fintech and banking platforms that serve tens of millions.

Distributed systems · Fintech & banking platforms · Kafka · Kubernetes · MCP · Local model pipelines
GitHub

What we're building, out in the open.

Products to use, code to read, writing to check.

Products
Open source
Writing & video
Let's build

Have something AI should be doing?

Tell us about the product, the workflow or the engineering problem, and what you have already tried. A software engineer reads it and replies within two business days.

Start a conversation