Distopik — Systems Where Knowledge Compounds

01 — Domains of Practice

What we practice

Most organizations don't have a knowledge problem. They have a retrieval problem. Three domains, one purpose: closing the gap between what your company knows and what it can use.

Knowledge Management

The connective tissue: ontologies, retrieval pipelines, and semantic infrastructure that turn scattered documents and departing experts into a system your teams — and your agents — can actually ask. The layer every other AI investment stands on.

Agentic Systems

Software that plans, calls tools, and completes work — with the eval harnesses, guardrails, and human checkpoints that make autonomy safe to run against production systems. Agents that earn their permissions.

AI & Machine Learning

Applied models in production — scoping, training, evaluation, and the unglamorous engineering that keeps them honest. We build ML that survives contact with real data, real latency budgets, and real users.

02 — Ways of Working

Two ways to engage

Every engagement is founder-led, senior only, and scoped in writing before work begins. The difference is where we stand.

Fixed Scope — For You

We take the problem off your desk.

A fixed scope, a schedule, and a handover date. We write the specification with you, build against it, evaluate it, and hand it over running: architecture, code, documentation, and the operating manual. You get a finished system, not a slide deck.

Embedded — With You

We work inside your team.

Embedded alongside your engineers for the messy middle: shaping architecture, pairing on the hard problems, reviewing what ships, and mentoring as we go. We leave when the capability is yours — and not before.

First Engagement — Fixed Scope

The Knowledge Infrastructure Assessment

Most engagements begin here. Over two to three weeks we inventory what your organization knows and where it lives, audit how well it retrieves — by people, by search, by agents — and deliver a build roadmap sequenced by risk and return.

The roadmap is yours either way — written to be executed by us, by your own team, or by the next firm you interview.

Book the assessment

Knowledge inventory — sources, owners, systems of record, and what lives only in people's heads
Retrieval audit — what your people, your search, and your tools can actually find
Agent-readiness review — APIs, permissions, and data quality against what autonomy requires
Build roadmap — architecture, sequencing, and estimates specific enough to contract against
Handover session — findings defended, hard questions answered

03 — Selected Work

Three engagements, annotated

Named where permitted, anonymized where discretion requires. The shape of the work is exact.

Knowledge Systems · Enterprise

An enterprise operating system on a knowledge graph

Problem: Wisdom Labs — the AI spinoff of a multinational grocery group — had a mandate to predict the business and optimize the decisions running it: stock, portfolio, the timing of replenishment. What it didn't have was the team or the architecture.
Built: As interim CTO: a 40-person data-science and engineering organization, recruited from zero on the engineering side — and the system itself, a Neo4j knowledge graph as the enterprise's shared memory, with a self-learning decision layer on the JVM above it.
Outcome: €250,000 in savings in a single quarter, in a single pilot location — dead stock and supply-chain slack, found and eliminated.

Custom Agents

An agent that scaled a one-founder coaching practice

Problem: A music-industry coaching company was selling like a tech product but running on one founder's personal judgment. Without automation — or a payroll full of coaches — it could not scale.
Built: As CTO: the product itself, an LLM agent embodying the founder's methods and source material, walking each artist through a customized yearly plan and weekly checklists, evaluating submitted work — and threading the human coach in exactly where judgment earns its keep. Classic ML mined years of client conversations and outcome data for what moved careers.
Outcome: The practice became a product: routine coaching runs as software, the founder's attention reserved for the moments that need them.

Knowledge Systems · Real-Time

Live speech to knowledge graph, three seconds behind the room

Problem: A live-events platform needed the spoken content of parallel stages to become structured, queryable knowledge while the event was still running — for organizers, attendees, and the products built on top.
Built: A provider-agnostic ASR front end feeding LLM claim, assertion, and entity extraction; disambiguation, ranking, and Wikidata linking; link prediction into a growing knowledge graph with epochal rollups — engineered for reliability across simultaneous stages.
Outcome: In production at its first paying customer, with pilot events converting to paid: real-time products hold a 3-second P95; full stage-to-graph refresh lands inside 15 minutes.

From the archive: the just-in-time imagery engine behind Sentinel Hub — petabyte-scale, ESA Copernicus Masters winner, acquired with Sinergise by Planet Labs · a boot-speed and performance rescue on a 100-engineer regulated-gaming RTOS codebase · a commerce SaaS taken international as interim CTO — React admin, public API with SDK, containerized core

Have a problem this shape? Book thirty minutes

04 — Expertise

An index of expertise

The practice rests on six load-bearing fields. They look unrelated until you need two of them at once. In our experience, every serious project does.

01

Knowledge Systems

Ontology and taxonomy design
Retrieval pipelines and semantic search
Knowledge graphs and entity resolution
Corpus ingestion, normalization, and permissions
RAG architecture and grounding
Expert-knowledge capture
Retrieval evaluation and relevance tuning

02

Custom Agents

Agent architecture and orchestration
Tool design and API surfaces
Guardrails and permission scoping
Human-in-the-loop checkpoints
Memory and context management
Eval harnesses and release gates
Observability and cost control in production

03

AI / ML / Data

System design and architecture
Model scoping, training, and fine-tuning
MLOps consulting
Data pipelines and data quality
Scaling and operations
Cloud infrastructure

04

Software Architecture

Architecture and design
Technology evaluation
Stack selection
Retooling and replatforming
Performance engineering and profiling
Scaling operations
DevOps and process automation

05

Testing & Evaluation

Test automation and CI integration
Complex systems testing
Performance and load testing
Data quality assurance
LLM and agent eval suites
Regression baselines for non-deterministic systems
Risk-based test strategy

06

Advisory & Mentoring

CTO-for-hire
Technical due diligence
Mentoring engineering leadership
Technical hiring and interview design
Build-vs-buy analysis
Vendor and model selection

05 — Correspondence

Start a working session.

Thirty minutes, no deck, no discovery-call theater. Bring the problem; we'll bring questions. If we're not the right practice for it, we'll say so — and usually name who is.

Current engagement concludes in October — booking Q4 now

Book thirty minutes Write to hello@distopik.com