Custom Software Development
Tailored web applications, internal dashboards, APIs, and cloud-based systems built around your real operations.
Solutions & Insights
Ideas for building intelligent products.
Explore what we build for businesses, then go deeper with practical writing on AI, software architecture, data, and product engineering.
What we solve
Solutions, explained.
The engineering behind our own products, applied to your business—from custom platforms and mobile applications to domain-specific AI and dependable data systems.
AI Agent Development
Enterprise AI agents that connect to your business systems, use tools and wait for human approval at critical steps.
LLM Training & Fine-Tuning
Domain-specific language models, Turkish NLP, fine-tuning, and adaptation to your enterprise data.
Computer Vision & Image Recognition
Object recognition, OCR, document analysis, and AI-powered vision systems for demanding environments.
RAG & Enterprise Semantic Search
Semantic search over corporate documents and grounded AI assistants that cite their sources.
Mobile App Development
Modern, responsive, AI-powered mobile products for iOS and Android, from design through release.
Big Data & Data Engineering
Reliable collection, processing, storage, and analytics infrastructure that makes data ready for AI.
Why Hatson Tech
Proven in products, applied to your project.
Our team brings its field-tested experience in Turkish AI, large language models, RAG, semantic search, software, and data engineering directly to your business.
Working across Türkiye
From the field
Clear thinking on complex technology.
Plain-language articles on large language models, RAG, AI agents, data privacy, Turkish NLP, and the decisions behind production-ready AI.
71 articles
Private Inference for Sensitive RAG
A private model may address residency, latency, customization, or control requirements. Includes a private rag data-path matrix.
5 min read
Email Triage Agent with Human Review
Build email triage that classifies and drafts with human review, measurable routing quality, and strict tool scope.
5 min read
OCR, Detection and Vision-Language Models
Select OCR, object detection, or a vision-language model from the exact output and evidence your workflow needs.
5 min read
Designing a Useful “No Answer” in RAG
Design a useful RAG no-answer path for missing, ambiguous, unauthorized, or contradictory evidence.
5 min read
Setting a Cost Budget per Agent Task
Set a per-task agent cost budget with step limits, tool quotas, graceful stopping, and outcome-based evaluation.
5 min read
Building RAG from Scanned PDFs
Build RAG from scanned PDFs with image-quality gates, OCR review, page provenance, and answer tests.
5 min read
Building an Agent Evaluation Dataset
AI agent evaluation with a task and failure case template: final state checks, a held-out dataset, repeated trials and release acceptance criteria.
8 min read
Diagnosing a Bad RAG Answer
RAG debugging: separate ingestion, permissions, retrieval, generation and citation failures with a fault tree, a worked example and regression tests.
9 min read
Prompt Injection Through Tool Output
What is prompt injection through tool output? Learn how to enforce agent permissions, validate approval and test for data leaks and forbidden actions.
9 min read
Human Approval Before Agent Actions
Human in the loop for AI agents: which actions need approval, why the gate belongs on the server, a refund example and how to measure review quality.
7 min read
When to Split a Monolith
When should a monolith become microservices? Compare service boundaries, data ownership and operating costs with a pricing example and rollback tests.
8 min read
Permission-Aware Retrieval for Enterprise RAG
Enforce source-system permissions inside enterprise RAG search: where to filter, what to store, a leakage test plan and common mistakes.
8 min read
Segmentation or Detection for Defective Parts?
Choose image segmentation or object detection from the reject rule; compare annotation work, area measurement and final part decisions in a pilot.
8 min read
Choosing Tools for an AI Agent
AI agent tool selection: build a narrow catalogue with clear contracts, server permissions, safe retries, a support example and realistic tests.
8 min read
A Strangler-Fig Migration Plan
What is the Strangler Fig pattern? Build a route-by-route legacy migration plan with data ownership, progressive rollout, cutover gates and rollback tests.
8 min read
RAG Over Versioned Policies and Procedures
RAG over versioned documents: choose the effective policy with date filters, passage metadata, a travel example and tests for change boundaries.
8 min read
Visual Anomaly Detection Without Defect Labels
Plan visual anomaly detection using normal images, independent defect tests, alert thresholds and operator workload before automating rejection.
8 min read
AI Agent or Fixed Workflow? A Decision Guide
AI agent or fixed workflow? Choose the scope with an autonomy decision tree, a support example, cost arithmetic and comparative acceptance tests.
8 min read
Few-Shot Vision for Defect Detection: Which Labels Do You Need?
Plan few-shot image recognition with labels per class, normal-only anomaly baselines, independent tests and an annotation budget for defect inspection.
8 min read
How to Verify Citations in RAG Answers
Verify RAG citations claim by claim: test passage support, version and access, then diagnose partial evidence with a worked policy example.
8 min read
Application Modernization: Rebuild, Refactor or Retain?
Application modernisation: choose retain, replatform, refactor or rebuild with a component decision matrix, worked cost example and migration test plan.
8 min read
In-House Enterprise AI Solutions: On-Premise AI
AI systems that work with your own documents, follow your access rights and show the source behind every answer: use cases, on-premise, private cloud and hybrid deployment, KVKK, and the path from pilot to production.
8 min read
An Ankara Software Company That Delivers Your Software and Stands Behind It
One team from analysis to maintenance for custom software, integration and mobile projects: how we work, the commercial model, source code ownership, reference projects and the practical case for a local partner in Ankara.
7 min read
The EU AI Act After the Omnibus: What Actually Changed in 2026
The Digital Omnibus on AI deferred Annex III high-risk obligations to 2 December 2027 and Annex I to 2 August 2028, while Article 50 transparency stayed on schedule. A clean timeline for engineering teams.
15 min read
Turkey's AI Regulation: Where It Actually Stands
No dedicated AI law is in force in Türkiye, but Law 6698, the Penal Code, Law 5651, the Commercial Code, sector regulators and EU customers already constrain what you can ship. The engineering answer.
14 min read
KVKK's Generative AI Guide: What Engineering Teams Must Change
KVKK's generative AI guidance turned into concrete architecture: logging boundaries, retention windows, redaction before inference, retrieval-level access control, shadow AI detection and erasure that reaches the vector index.
15 min read
AI Governance Engineers Actually Run
Checklist-grade AI governance: the fields an inventory row must carry, a tiering rubric a team can score in twenty minutes, three oversight patterns with their real costs, change control and an incident taxonomy.
14 min read
What an AI Project Actually Costs
AI vendors quote ranges that share no definition of what is inside them. Here is the cost structure instead: six budget lines with typical shares, a worked 45 person-week example, and why a fixed price is a warning sign.
15 min read
Why AI Pilots Never Reach Production
Most AI pilots do not fail. They just never end. Six concrete failure modes between the demo and production, what Gartner's June 2025 press release actually found, and the exit criteria a pilot must pass to earn budget.
15 min read
Build vs Buy for AI: An Honest Decision Framework
Every vendor page answers buy and every dev shop answers build. Six real decision axes, three-year cost maths including per-seat escalation, a scored rubric you can fill in, and the branches where we tell clients to buy.
15 min read
How to Choose an AI Development Partner
The vendor evaluation instrument nobody publishes, because a fair one disqualifies its own author. What to ask about production evidence, IP and prompts, exit clauses, red flags, and how to run a two-vendor bake-off.
15 min read
Data Readiness: The Assessment to Run Before Any AI Project
A practical data readiness instrument: nine dimensions scored zero to three for 27 points, with observable rubrics, remediation effort bands, veto rules, and how to cut a slice that is good enough for the first workflow.
14 min read
Measuring AI ROI Without Fooling Yourself
Hours saved are not money. An engineer’s guide to AI ROI measurement: pre-launch baselines, cost per workflow execution, deflected versus realised cost, holdout design, payback arithmetic and kill criteria.
14 min read
How to Scope an AI Project So It Can Actually Ship
A one-page scoping template with sixteen fields: a testable task definition, an acceptance number set against the measured human baseline, an evaluation set frozen during scoping, and a kill gate somebody can actually pull.
14 min read
How to Tell If Your RAG System Is Any Good
Recall@k, context precision, faithfulness, citation correctness and refusal behaviour, plus how to build a golden dataset, calibrate an LLM judge and gate regressions in CI before they reach users.
15 min read
Chunking at Scale: Where Retrieval Quality Is Won or Lost
Layout-aware parsing, OCR quality gates, four chunking strategies and when each wins, size and overlap as tunables, metadata that outperforms the embedding, and re-indexing millions of documents without a full rebuild.
15 min read
Hybrid Search and Reranking: Beyond Pure Vector Similarity
BM25 and dense retrieval fused with reciprocal rank fusion or weighted scores, cross-encoder reranking over wide candidate sets, query rewriting, pre-filtering for access control, and HNSW recall versus latency.
15 min read
Evals and Observability: Running AI After Launch
Traces with prompt and model versions, pre-merge eval gates, judge calibration and drift, the three kinds of model drift and how to detect each, AI incident response, and the compliance evidence it all produces.
15 min read
Prompt Injection, Data Exfiltration and Guardrails
Direct versus indirect injection, the four channels data actually leaves through, excessive agency, what guardrails can and cannot do, and the RAG-specific controls that make retrieval safe to run in production.
15 min read
Model Routing: The Biggest Cost Lever Most Teams Haven't Pulled
Where LLM spend actually goes, how to build a routing layer, the four router types and their latency cost, cascades, prompt caching, and the pre-merge eval gate that catches silent quality regression.
15 min read
Long Context or RAG? The 2026 Answer
What actually changed in long context, why context rot and position effects still matter, the real cost arithmetic, when each approach wins, and the hybrid that most production systems converge on.
15 min read
The Economics of Fine-Tuning: When It Pays and When It Doesn't
The four things fine-tuning buys and the one it does not, why dataset construction dominates the bill, LoRA versus full training versus distillation, and a worked break-even by request volume.
15 min read
On-Prem and Sovereign AI for Turkish Enterprises
Sovereign-AI guides cover the US, UK and Gulf. This one covers Türkiye: residency versus sovereignty, the three deployment postures, KVKK Article 9 transfers, and the utilisation number that decides whether kurum içi pays.
14 min read
Agentic Workflows in Production: What Actually Breaks
What actually breaks when AI agents reach production: unbounded loops and the bills they generate, tool errors the agent hides, permission scope creep, and the control set that keeps a live agent survivable.
15 min read
GEO: How to Get Your Company Cited by AI Search
Static HTML, reciprocal hreflang, BlogPosting and FAQPage schema, named allow blocks for AI crawlers and a hand-written llms.txt: what we implemented on hatsontech.com for GEO, and what we deliberately skipped.
14 min read
Computer Vision for Quality Control: The Engineering Reality
What actually decides a production quality-control vision project: lighting and fixturing, how much labelled data you really need, why accuracy is meaningless at a 0.5% defect rate, and setting the threshold from cost.
16 min read
AI in Finance: Where It Works and Where It's Regulated
What AI genuinely does in banking today: document extraction, contract review, policy answers with citations. Plus the Annex III high-risk regime, its new 2 December 2027 date, KVKK article 11(g), and an audit-ready architecture.
14 min read
Reasoning AI Models: How 'Thinking' Models Actually Work
A new class of models does not answer right away; it 'thinks' first. Here is what that thinking really is, why it solves certain problems, what it costs, and why you do not need it for every task — without hiding behind jargon.
12 min read
What Is MCP (Model Context Protocol)? The Standard Connecting AI to Tools
How MCP hosts, clients, and servers connect tools and data to AI applications, with a practical example, security checks, and guidance on when a direct API is simpler.
5 min read
Context Engineering: Beyond Prompt Engineering
Getting good results is no longer a single 'right question'; it is designing the entire context the model sees. Instructions, retrieved documents, memory, tool outputs — assembling them well is called context engineering.
5 min read
Small Language Models (SLMs) and Edge AI: Is Small Always Weak?
How to choose a smaller model with real task tests, hardware measurements, privacy checks, and an escalation path for uncertain requests.
5 min read
Multi-Agent Systems: From a Single Agent to an AI Team
When independent work benefits from several agents, when one agent is enough, and how to compare quality, source support, latency, and cost.
5 min read
What Are Large Language Models (LLMs) and How Do They Work?
Understanding the technology behind ChatGPT, Gemini and Claude from scratch: from next-token prediction to parameters, from context windows to the difference between training and inference, with no technical background required.
12 min read
Tokens, Embeddings and Vectors: How AI Turns Language Into Numbers
A computer processes numbers, not letters. So how does a word, a sentence, even a meaning turn into a number? From tokenization to embeddings, from vector space to semantic similarity, we explain how AI grasps language with concrete analogies and no math required.
11 min read
Prompt Engineering: A Practical Guide to Getting Better Answers from AI
The same AI model can hand you a forgettable answer from a sloppy prompt and a genuinely useful one from a good prompt. We unpack the few simple but powerful habits behind that gap —clear instructions, context, examples, and step-by-step thinking— with concrete before-and-after comparisons.
11 min read
RAG or Fine-Tuning? Which to Use and When
There are two distinct ways to give a language model what it needs to know: leave the book open in front of it (RAG) or send it back to school (fine-tuning). Here is what each one really does, where they diverge, and why the best systems often use both.
11 min read
AI Hallucination: Why Models Make Things Up and How to Reduce It
Why does a language model state a wrong answer with total confidence, as if it were true? In plain language, we explain what hallucination is, why it springs from the model's very nature, and how engineers rein it in.
11 min read
The Transformer Architecture, Explained Simply
A single paper published in 2017 changed the course of AI. We walk through why the "Attention Is All You Need" idea matters so much, how words learn to "pay attention" to one another, and why nearly every large language model today rests on this one foundation.
11 min read
AI Agents and Tool Use: From Chatbot to AI That Actually Does Things
What separates an "agent" from a plain chatbot? We explain the plan-act-observe loop, tool calling, MCP-style tool standards, and where agents shine and where they fall apart, in plain language.
12 min read
Vector Databases and Semantic Search: How a Computer Finds "Meaning"
From keyword search to semantic search: embeddings, ANN indexes like HNSW, what a vector database actually does, and re-ranking — all explained with plain examples, plus how an AI assistant finds documents by meaning rather than by words.
12 min read
AI Ethics, Bias, and Safety: A Calm, Practical Guide
Where does bias come from, what do fairness and transparency really mean, and what do KVKK and the EU AI Act actually ask of us? A concrete walkthrough, without the lecture.
11 min read
AI in Law: Opportunities, Risks, and Responsibility
AI is collapsing legal research, drafting, and review from hours into minutes. But it also brings real risks: hallucinated citations, confidentiality leaks, and a quiet creep of over-reliance. This piece explains what genuinely works, what is dangerous, and why the final word still belongs to the lawyer.
11 min read
AI at Work: A Calm, Practical Productivity Guide
How professionals use AI every day, from writing to research to email to code, how to delegate to it well, and exactly where you should not trust it.
11 min read
Multimodal AI: How One Model Understands Text, Images and Sound at Once
When an AI can look at a photo and read the text inside it, interpret a chart, or listen to a voice and respond, we call it multimodal AI. We unpack the idea beneath this ability, its real-world uses and its limits, with concrete examples and no technical background required.
12 min read
How Do We Know an AI Model Is Good? Benchmarks and Evaluation
What those leaderboard percentages really mean, where benchmarks mislead, and what actually proves a model works in the real world. A plain-language guide.
12 min read
Open Source or Closed Model? AI's Big Trade-off, Explained
An AI model is either something you rent like a cloud service or something you download like a file and run on your own machine. Choosing between the two is a real trade-off among control, privacy, cost, and capability. Here is what each one means, when each makes more sense, and why mature teams usually run both.
12 min read
AI and Data Privacy: Where Does Everything You Type Actually Go?
When you type something into an AI tool, where exactly does your data go? The critical difference between training and inference, the cloud–private–on-prem spectrum, the basics of KVKK and GDPR, the real limits of anonymization, and practical tips that actually work for both individuals and companies.
12 min read
AI for Businesses: Where to Start Without the Hype
A practical roadmap for companies that start with a real problem instead of a slide full of promises — that go small, measure honestly, and know when to walk away.
11 min read
The Future of AI and the AGI Debate: Separating Signal from Hype
What does "AGI" actually mean, where are we really right now, and what does agentic AI change? We calmly separate the optimistic and cautious views, the genuine near-term impact, and the inflated promises.
13 min read
Grounding LLMs in Your Business Data with RAG
Why a raw language model gets domain-specific facts wrong, how Retrieval-Augmented Generation fixes it, and how we build a retrieval layer that grounds answers in your own data.
12 min read
AI for Turkish: The Challenges and Opportunities of Turkish NLP
Why Turkish is hard for language models, how that difficulty shapes LLM behavior, and what it really takes to build Turkish-first AI products.
11 min read
No articles match that search yet.
Have a real problem to solve?
Start a project