Agent architectures
How to give a model tools, memory and a task boundary it won't wander out of — and how to tell when it has. Duckerhub is where we test this against real users.
Applied AI lab · Bratislava
We do the research.
Then we ship it.
Sinorom is an applied AI lab. We work on language models, agent architectures and the data infrastructure underneath them — and everything we conclude gets tested the only way that counts, by running in production under real load.
- Published & patentedPreprints, articles and granted US patents
- We run what we buildFive products live on our own infrastructure
- Evaluated, not demoedMeasured on your cases before it ships
Research
What we're working on.
We keep a handful of open questions running at any time. They get picked because a client problem hit a wall, or because we needed the answer for one of our own products — never for their own sake. Everything here is applied: if it doesn't eventually change what we ship, we drop it.
Grounding & retrieval
Getting the right context in front of the model: chunking, ranking, and the underrated problem of recognising when the answer simply isn't in the corpus.
Evaluation methodology
Most AI projects stall because nobody agreed what “better” means. We build the measurement before the feature, and keep it running afterwards as a regression suite.
Database internals
Separating database compute from durable logs and immutable object storage. DBlicious is the long-running experiment, and the hardest thing on our bench.
Services
What the lab does for clients.
The research is the interesting part; this is how it reaches you. Most engagements start in one of these and grow into another — we're happy to own the whole lifecycle, or just the part your team can't cover right now.
AI systems & agents
Models and agents taken from prototype to something you can depend on.
This is the centre of what we do. We build AI systems that survive contact with production: grounded in your own data, measured against real cases before and after the change, and with defined behaviour for the times the model is wrong. Where the received wisdom doesn't fit your problem, we run the experiment ourselves rather than guessing.
- Retrieval-augmented generation and grounding
- Agent design, tool use and multi-step workflows
- Evaluation harnesses and regression suites
- Document processing, extraction and classification
- Prompt and context engineering, versioned like code
- Model selection, routing and cost/latency tuning
- Fine-tuning and small-model deployment
- Data pipelines and warehousing to feed them
- Self-hosted and EU-resident inference
Typical engagement — a discovery spike that establishes a baseline and an evaluation set, then a scoped build once the numbers justify it.
Custom software development
Products, platforms and internal tools — built to spec and built to last.
Most of our work begins with a business problem rather than a finished specification. We scope it properly, agree what “done” means in writing, and then build the thing: a real application with tests, documentation and a deployment story — not a prototype you have to throw away in six months.
- Web applications and internal platforms
- APIs, integrations and third-party connectors
- Legacy replacement and incremental migration
- Database design, modelling and performance work
- Automated testing and CI pipelines
- Documentation and clean technical handover
Typical engagement — a fixed-scope project, or a continuous build on a monthly retainer. Code lives in your repository from the first commit.
Cloud, DevOps & infrastructure
The unglamorous half of shipping software, done without shortcuts.
Software that isn't deployed, monitored and backed up isn't finished. We build the infrastructure your application actually needs — no more, no less — automate the parts that should never be manual, and leave you with a system your own people can operate without calling us first.
- Linux provisioning, hardening and patching
- Docker and container orchestration
- CI/CD pipelines on GitLab CI or GitHub Actions
- Nginx, TLS certificates and reverse proxying
- Monitoring, alerting and log aggregation
- Backups, restore tests and recovery drills
- Infrastructure as code and repeatable rebuilds
- Hosting cost review and right-sizing
Typical engagement — a one-off build with full handover, or ongoing managed operations with an agreed response window.
Consulting & technical partnership
A senior engineer in the room when the decision is expensive.
Sometimes you don't need another pair of hands — you need someone to tell you whether the plan is sound. We review architectures, sanity-check vendor proposals, and take on longer-term advisory work for teams that have outgrown decisions made on the fly.
- Architecture and code review
- Technical due diligence before an investment or acquisition
- Fractional CTO and advisory retainers
- Vendor, platform and build-versus-buy selection
- Mentoring and review practice for in-house teams
- Security posture and dependency audits
Typical engagement — a half-day review with a written summary, or a recurring monthly retainer with a standing call.
Work
Where the research gets tested.
Client work is mostly under NDA, so here is the better evidence anyway: our own products, in production, on infrastructure we operate. Each one exists because a question needed answering at real scale rather than in a notebook.
Duckerhub
LivePocketBase backends on demand — database, auth, file storage and realtime on your own subdomain. Sleeps when idle, wakes on request.
Jipidy Grid
LiveCross-browser visual testing. Screenshots, visual regression and scripted flows on Chromium, WebKit and Firefox — desktop to phone — from one API call.
Golem
LiveGives communities the machinery to reward their own people — leveling, moderation, AI chat, and a member economy where time spent turns into value members actually hold.
eBVB
LiveA market-intelligence platform for the Bucharest Stock Exchange — real-time quotes, order-book depth and a screener. Romanian-language.
Skimetric
LiveEuropean ski intelligence across 1,577 resorts — verified snow forecasts, live lift and piste status, ten-winter history, and AI trip planning.
DBlicious
In developmentPostgreSQL and MySQL, unbundled — upstream database compute separated from durable write-ahead logs and immutable S3-compatible storage. Our current deep-infrastructure project.
Approach
How an engagement actually runs.
No discovery theatre and no open-ended hourly drift. Four stages, and you can stop after any of them.
Understand
We start with the problem, the constraints and the deadline that genuinely matters. If we think you're solving the wrong thing, you'll hear it before you spend money.
Scope
You get a written proposal: what's included, what deliberately isn't, what it costs and when it lands. Changes are re-quoted rather than absorbed silently.
Build
Short cycles, working software you can click on, and a direct line to the person writing the code. Everything sits in your repository from day one.
Operate
We deploy it, watch it, and hand over documentation someone else could act on. If you'd rather we kept running it, that's a separate, honest line item.
Stack
Tools we reach for.
Boring, well-understood technology by default. We'll happily work in yours instead — the list below is preference, not a requirement.
Models & inference
ClaudeOpenAIOpen-weight modelsvLLMOllama
Retrieval & evaluation
pgvectorHybrid searchRerankersEval harnessesTracing
Languages
PythonTypeScriptGoSQLBash
Backend & data
Node.jsFastAPIPostgreSQLRedisAirflowREST
Frontend
ReactNext.jsVueTailwind CSSVite
Infrastructure
LinuxDockerNginxTerraformGitLab CIS3Hetzner
About
A lab small enough to ship.
Sinorom is deliberately small, and there is no wall inside it between the people who try things and the people who make them run. The same engineers write the evaluation harness, argue about the results, and carry the thing to production — which is the only arrangement we've found that stops research becoming a slide deck.
We fund the open questions by building products out of the answers. Duckerhub, Jipidy, Golem, eBVB, Skimetric and DBlicious are all places where an idea met real users and real load before we ever recommended it to a client.
We work with founders, product teams and IT departments around the world — usually remotely, and used to running a project across time zones. We're comfortable being your only technical partner, and equally comfortable slotting in alongside a team that already knows what it's doing.
“Measure it before you believe it. Then measure it again in production, where it counts.” — the only house rule
Publications & patents
The work, in the open.
Co-authored US patents from work at IBM and Kyndryl.
All on Google Patents-
Data protection and recovery
Generates a backup and recovery option by classifying an application's criticality from its static and runtime metadata using machine learning. Kyndryl · granted Jan 2023
-
Client-less database system recovery
A trigger on the primary database normalises every data modification into a universal format that any of several backup databases can consume. IBM · granted Jan 2023
-
Tracking a potential attacker on an external computer system
Categorises a connecting external system as a potential threat and tracks it from the requests it makes. IBM · granted Oct 2025
-
Adaptive network control of traffic flows in a secure network
Rate limiting driven by a supervised-learning model that matches each traffic flow against pretrained flow models. IBM · published Jun 2023
-
Data synchronization in edge computing networks
Synchronises data across an edge network by electing a leader node, with follower nodes each casting a single vote for a candidate. IBM · published Nov 2022
Preprints, articles and theses by Marcel Butucea.
All 12 on ResearchGate-
Make Attention Sub-Quadratic Again
Learned ANN-compatible search projections for post-hoc sparse attention.
-
Certified Block Screening for Adaptive Long-Context Attention
Screening attention blocks with a correctness certificate, for long-context inference.
-
From “Truth Topology” to Readout Bottlenecks
Negative replication on GSM8K, and a positive representation–decoder dissociation in procedural micro-worlds.
-
Scoped Actions Expose Compositional Failures in Verified Symbolic Rewrite Search
Where multi-step symbolic search breaks down once actions are given scope.
-
RegimeLift
Targeted residual uplift across nested Gemma 3n operating regimes.
-
The Obsolescence of Effort
Software engineering metrics, theory of constraints, and the decoupling of human effort from software output.
-
From Shreds to Fills
Solana launch microstructure, bonding-curve graduation and execution realizability — a month of full-firehose observation and failure-aware experiments.
Contact
Tell us what you're trying to build.
A short description of the problem is plenty to start. We reply to every genuine enquiry, usually within one business day.
General enquiries
contact@sinorom.roNot sure who to ask for? Start here and we'll route it internally.
New projects
sales@sinorom.roScoping, proposals, timelines and pricing for new work.
Client support
support@sinorom.roFor clients with an active project or a support agreement in place.
Careers
careers@sinorom.roOpen applications are welcome, even when nothing is advertised.
Accounts & invoicing
billing@sinorom.roInvoices, purchase orders and VAT questions.
Privacy & GDPR
privacy@sinorom.roData protection requests and questions about how we process data.
Correspondence address: Karpatské námestie 10A, 831 06 Bratislava, Slovak Republic.