Hi, I am Konstantin Knyazkov


I build intelligent systems
for complex real-world problems — from problem framing and research to production.



Technical Lead AI & Software Engineer Applied Researcher Systems Architect Entrepreneur

PhD in Computational Science · 15+ years across industry and research · Based in France

I can help with

I work across business, research and engineering on ambiguous, technically difficult problems — using AI where it creates real leverage, from understanding the problem with clients and domain experts, research if needed, to building the production system.

Applied AI & LLM systems Data-intensive applications Complex systems & simulation Optimization R&D prototypes Software architecture

Where I add most value

Financial institutions & risk teams

Computational risk, quantitative workflows, distributed calculation and analytical platforms.

Deep-tech & AI startups

Hard technical problems, PoCs, architecture, AI-enabled R&D and research-to-production.

Scientific & engineering companies

Simulation, optimization, HPC and computation-intensive R&D pipelines.

Enterprises applying AI

Complex data, legacy systems, expert workflows and AI-enabled analytical products.

Selected work

Everix — Risk Platform

Co-founder · Production · Distributed quantitative systems

Production computational risk platform for investment banks and funds, delivered as SaaS and on-premise software.

Selected problems

Product website everix.io ↗

Enterprise AI

Technical lead · LLM analytics · Enterprise data

LLM-driven analytics over large legacy enterprise databases, combining semantic metadata, Text2SQL, validation, agentic workflows and visualization.

Selected problems

Scientific computing

Production · Published work · Distributed systems

Research and engineering across distributed scientific workflows, computational steering and large-scale simulation.

Selected problems

Selected publication ↗ Product demo ↗

Healthcare transport & decision support

Published research · Medical collaboration

Research with medical partners on time-critical patient transportation, hospital accessibility and complex clinical data.

Selected problems

Selected publication ↗

Format of intervention

I work on a B2B contract basis, from focused technical projects to fractional leadership, with each engagement structured around a concrete outcome — a decision, prototype, architecture or working system.

Short technical / R&D project

Frame an ambiguous business or technical problem, work with stakeholders, and use AI, research and engineering to determine and build the right solution.

Several days to ~3 months

Build / unblock

Design and implement a difficult subsystem, resolve architectural uncertainty, or get a stalled effort moving.

Several weeks to ~3 months

Fractional technical leadership

Architecture, technical decisions, reviews, coordination and hands-on implementation where useful.

Typically 1–3 days/week

Flexible involvement · Controlled budget · Senior capacity when needed

Background

These are selected examples rather than a complete project history. For the broader career and research record:

Contact

Have a difficult technical problem?

Let's talk →

Everix · Incremental computation

How to compute a What-If scenario in seconds when a full calculation can take six hours?

A risk-management run could span ETL, pricing, sensitivities, SA-CCR, SIMM, xVA and corresponding aggregations. Repeating the whole graph for a narrow change such as adding one trade would make pre-trade analysis impractical.

I designed a scenario engine that starts from a completed parent run, discovers which dependencies are invalidated, recomputes only the affected subgraph and incrementally updates aggregation while reusing unaffected results.

Result: a full portfolio calculation could take ~6 hours; a narrow add-trade what-if could return in seconds.

Everix · Streaming aggregation

How to aggregate huge Monte-Carlo pricing results without loading hundreds of GBs into memory?

Monte Carlo xVA workflows could produce heavy pricing grids at trade-leg level, with representative shapes around 10,000 paths × 100 time steps. Retaining all child grids until pricing finished would create severe memory pressure.

I designed hierarchical streaming aggregation: partial results are reduced as they arrive, a parent is propagated upward as soon as all its children are complete, and heavy child data is released immediately. Pricing and aggregation therefore run as one pipeline instead of two sequential phases.

Result: lower peak memory and shorter end-to-end latency by overlapping pricing and aggregation.

Enterprise AI · Semantic data

How to make AI reason over a database whose schema does not contain enough information to understand the data?

The source environment contained hundreds of legacy tables. Names were cryptic, descriptions missing, joins implicit and dictionary semantics known only to experienced internal analysts. Raw DDL was therefore not enough for reliable Text2SQL.

I designed a semantic metadata layer and onboarding workflow combining database metadata, expert descriptions, hidden relationships, value domains and content indexing. This semantic representation then fed schema selection, SQL generation and validation.

Result: deployed on real enterprise data; large tables could contain ~120 columns.

Enterprise AI · Human + AI analytics

How to understand how employees actually use an AI assistant from 70k+ free-form interactions?

A single classification axis was not enough: “what the user wants the AI to do” and “what business topic the user is working on” are different dimensions. I designed a hierarchical two-axis taxonomy and used an LLM to bootstrap candidate classes from a sample of requests.

The taxonomy was manually consolidated, then classification was run with correction stages to prevent invented or inconsistent labels. This turned conversational logs into structured product and organizational analytics.

Result: tool to analyze current and historical usage of the system, started road to unsupervised use-case detection.

Scientific computing · Interactive workflows

How to turn batch scientific workflows into interactive computational experiments?

Traditional scientific workflows assume that components run as batch jobs and exchange data only between stages. That model breaks down when a user needs to steer a running simulation, stream intermediate state or modify an experiment at runtime.

I worked on an extended workflow model with long-running components, runtime communication, events and commands, streamed data and dynamic workflow changes. The same abstraction could represent software, data sources, visualization components and external devices.

Result: published work, implemented in production within a eScience platform.

Scientific computing · Urban simulation

How to orchestrate arbitrary scientific software across desktops, clusters and supercomputers before containers existed and cloud was not mainstream?

I implemented the core abstractions that made this possible in the production platform: the computational package description model, workflow DSL and semantics, interpreter, and parts of the event-driven execution architecture. Once software was described as a package, domain researchers could compose it into workflows and execute experiments across heterogeneous infrastructure without dealing directly with deployment and orchestration details.

The production platform was used by multiple research groups across domains including quantum chemistry, hydrometeorology, transport simulation and computational sociology.

Result: owning core components of the production platform

Healthcare transport · Dynamic routing

How to choose the right hospital for a cardiac emergency in a large city when traffic conditions keep changing?

For acute coronary syndrome, geographic assignment alone can become suboptimal as traffic changes. Working with the Almazov Medical Centre, I combined road topology, population, hospital and traffic data to study how the best receiving hospital changes during the day.

I implemented OSM ETL and cleaning, accessibility analysis, visualization and the analytical logic used to compare stable versus dynamic destination strategies.

Evidence: published results with quantitative analysis showing advantages dynamic routing, prototype tool for ambulance drivers

Healthcare data · Legacy systems

How to extract clinical research data from a legacy system nobody can tell you how to query?

A hospital research project depended on data stored in a closed custom information system with poor documentation, no suitable API and no readily available extraction expertise. The research deadline could not wait for a perfect integration.

I coordinated access with the medical centre and used UI automation as a tactical extraction path to unblock the deadline, then moved toward a normal automated process. I also covered data quality layer, cleaning terminology and visualizing longitudinal patient trajectories.

Evidence: research-critical data was delivered in time.