AI

Invention
Lab

Research-grade solutions to AI problems
that engineering alone can't solve

Contact us

We are scientists specializing in optimization theory - how to extract maximum performance from AI
under real-world constraints.

Clients

Services

Products

Infiano Vision

Computer Vision

Industrial computer vision at research precision

Standard vision models are built for general cases. Infiano Vision is built for the cases where general isn't good enough — defect detection at 0.5 color tone deviation, dimensional accuracy at 1mm, clothing size estimation from a phone photo at 95% accuracy.

The system combines multiple specialized models into a single pipeline, calibrated for each deployment environment. Used in manufacturing quality control, retail, and fashion.

Go to Infiano Vision site

AI Tuner

LLM

Automated configuration for multi-agent AI systems

Multi-agent AI pipelines have hundreds of variables — which model handles which task, how prompts are structured, how outputs chain together. Teams spend weeks tuning these by hand, with no systematic way to know when they've found a good configuration.

AI Tuner analyzes real interaction data and finds the optimal setup automatically. The result is a better-performing system, reached in a fraction of the time.

Try AI Tuner

AMM Finance

Data Analysis

A decentralized exchange where liquidity is always in the market

The algorithm managing liquidity on every major DEX hasn't changed since 2018. At any given moment, 45% of LP capital sits outside the active trading zone and earns nothing — the only fix is constant manual monitoring.

AMM Finance solves this at the algorithm level. The AI tracks price in real time, keeps liquidity in the active zone automatically, and adjusts fees to volatility. Analysis of a real Uniswap pool confirms the approach: optimized positioning raises LP returns from 14.3% to an estimated 26%.

Go to Amm.finance

Optimization Theory in AI

Our research area

Optimization theory in machine learning is our primary field of study. We develop new methods, file patents, and publish findings — work that happens independently of any client or product.

6Patents
filed
2Patents pending
2Scientific publications

The discipline is about one question: how do you get the best possible result from an AI system when real constraints stand in the way — limited data, compute budgets, accuracy requirements, privacy rules. We have been working inside these constraints long enough to develop original answers.

Client projects and products are where that research meets practice. When we take on a problem, it goes through the same scientific process as our independent work.

Research Areas & Cases

Training Optimization

How to train a model faster, on less data, with guaranteed improvement at every iteration — without throwing more compute at the problem.

All cases

Company

Infiano is an AI invention lab based in Paphos, Cyprus.
We were founded with a specific conviction: that the most important AI problems in business are not integration problems, they are research problems and they require a scientific approach, not a delivery pipeline.

Our team combines academic researchers and engineering architects. We hold 6 patents across computer vision, language models, sound processing, and data analysis, with peer-reviewed publications at top machine learning conferences.

Every client engagement runs through the same process as our independent research. We study the constraint, develop or apply the right method, and ship a solution with measurable accuracy guarantees.

Ernesto Evgeniy Sanches Shayda

Ernesto Evgeniy
Sanches Shayda

Founder & AI Research Lead

Ernesto spent seven years as a researcher at Samsung Electronics, working on AI systems at industrial scale. He then pursued doctoral research in computer science at the University of Haifa under Professor Dan Feldman, one of the leading researchers in data summarization and algorithmic machine learning.

He founded Infiano to close the gap between research-grade methods and real business applications: to take the kind of work that gets published at top conferences and make it available to companies solving hard operational problems.

Contact with Ernesto

Why Infiano

Original research, when existing methods fall short

Most AI providers wrap existing models or fine-tune what's available. When a problem requires something that doesn't exist yet, they stop. We go back to the scientific question and develop a method for the specific constraint.

Documented
accuracy

Our results are tied to methods we own: under 1% error in distance measurement, 95% in clothing sizing, defect detection at 0.5 color tone deviation. These numbers are reproducible because they come from the method, not the benchmark.

Knowledge that transfers across domains

Research for manufacturing precision has improved our medical imaging work. Methods from financial analysis informed our DeFi algorithm. Every engagement draws on findings - clients get more than the solution to their problem.

Background IP - research depth without research cost

We retain the general methods we develop, clients receive the production implementation. This lets us invest in deep research independently, which means the science behind your solution keeps improving after delivery.

Academic and industrial depth in one team

Seven years at Samsung. NeurIPS publication. Six patents. Most teams operate in one register - either applied engineering or academic research. Hard problems require both.

How we work

We work with startups that need a technical team, enterprises looking for an embedded R&D unit, and AI teams that need their product to perform better. The process is the same in each case.

1

Problem Diagnosis

We begin by understanding the constraint — what standard methods deliver on your problem, where they fall short, and why. Not every problem requires original research. If an existing method works, we use it. If it doesn't, we develop one.

2

Research
& Development

Once the constraint is defined, we apply or build the right approach. This phase runs the same way as our independent research: methodically, with clear accuracy targets agreed upfront. Progress is measurable, not estimated.

3

Delivery

We ship production-ready systems with documented accuracy. Where the work produces new methods, those are filed as intellectual property under our Background IP model. Clients receive the implementation; we retain the underlying research.

Write to us

Industries

Constraints exist in every industry. The optimization methods we develop in one domain transfer to another — because the underlying problem is the same.

Retail & Fashion

Clothing fit measurement from a phone photo at 95% accuracy. Real-time detection of items added or removed from a cart. Consumer product analysis at a depth standard vision models can't reach.

Industrial
Manufacturing

Defect detection at 0.5 color tone and 1mm dimensional deviation using standard cameras. Vision systems for production lines where off-the-shelf models aren't precise enough.

Healthcare
& Medtech

Medical imaging analysis that catches what specialists miss. Our MRI models detect pathologies with accuracy exceeding routine clinical review — for diagnostics where false negatives have consequences.

Finance

Custom models for time series analysis and market pattern recognition. We identify structural similarities between historical market periods to support prediction without generic forecasting tools.

Education
& Edtech

Anti-cheating systems that analyze camera and microphone simultaneously. Diction assessment for HR interviews. Physical robot teachers for remote classrooms. AI that handles the parts of education that require genuine precision.

Documents
& Engineering

Recognition of complex technical documents — formulas, blueprints, machine schematics — where standard OCR fails. Custom models for document types that require domain understanding.

All industries

Publications

We write about the research behind our work: methods, findings,
and the problems that made them necessary.

September 15Machine Learning

Why your AI model doesn't need more data — it needs better data

Ernesto Sanches

Most teams facing an accuracy problem respond by collecting more training data. In many cases, that's the wrong move. We look at what the research on data compression says about which examples actually move the needle — and what to do with the rest.

Full article
July 25LLM

The hidden cost of tuning multi-agent AI by hand

Christian Johnson

When a multi-agent AI system underperforms, the usual fix is manual configuration: adjust the prompts, swap the model, try again. This works until the system is complex enough that the search space becomes intractable. We built AI Tuner to solve this systematically — here's what we learned in the process.

Full article
August 11Big Data

What a coreset is, and why it matters for enterprise AI

Johanna Nix

Coresets are a mathematical tool for reducing a dataset to its most informative elements without losing the properties that matter for training. They were developed in theoretical computer science, and they have direct practical applications for anyone training models under compute or data constraints.

Full article
September 4AI research

The point where scaling
stops working

Lius Quintero

More compute improves AI performance up to a point. Past that point, the bottleneck isn't resources — it's the method. This is the boundary where engineering alone runs out and research begins. How to identify it, and what to do when you're there.

Full article
August 26Computer Vision

How we measured distance with under 1% error using a standard camera

Kimi Chang

The industry standard for camera-based distance measurement sits around 10% error. We needed under 1%. This is a walkthrough of the problem, what we tried, why standard approaches failed, and the method we developed and patented.

Full article
All publications

Contacts

infiano Write to us

Gladstonos 12-14, 8046 Paphos, Cyprus