Insights

Engineering &
AI insights

What we are shipping and evaluating this year — AI systems, .NET, search, and the production constraints behind them.

Sep 11, 2026 Umut Kaya

Voice Agents in Production: Latency, Interrupts, and Trust

How Gidysoft ships voice agents: time-to-first-audio budgets, barge-in, honest fallbacks, and why we still split ASR, policy, and TTS in Polylingo.

Sep 08, 2026 Ayşe Demir

Hybrid Search for Learning Products: Why Vectors Alone Failed Us

Why Gidysoft dropped embeddings-only RAG in Polylingo: filters, BM25 plus vectors, lesson-sized chunks, and an index that updates when a teacher publishes.

Sep 02, 2026 İdris Yıldız

LLM Evals Before Launch: A Studio Checklist That Caught Real Bugs

The eval checklist Gidysoft uses before shipping LLM features: schema, grounding, refusals, live-trace sampling — and the Polylingo bug a spreadsheet caught.

Aug 26, 2026 Merve Yavuzer

Open-Weight vs Hosted Models: How We Choose Per Product

Gidysoft’s per-product choice between open-weight and hosted LLMs: data class, cost, eval floors, version pinning, and a .NET sidecar — not ideology.

Aug 12, 2026 Volkan Karataş

Multimodal Product UX: Vision and Voice That Earn Their Place

When Gidysoft adds camera and voice to products: task-first multimodal UX, latency, speakable errors, and a text fallback for real classrooms.

Jul 29, 2026 Zeynep Keskin

Prompt Injection When Your Agent Can Call APIs

How Gidysoft treats prompt injection as a confused-deputy bug: typed untrusted text, C# policy around tools, and CI fixtures — not prompt filters alone.

Jul 15, 2026 Umut Kaya

Semantic Caching: Cut LLM Cost Without Making Answers Stale

Gidysoft’s semantic cache for LLM calls: similarity thresholds, prompt-template invalidation, Redis plus vectors, and when we turn the cache off.

Jul 01, 2026 Mustafa Oğuz Demirel

The EU AI Act as Engineering Work, Not a Legal PDF

How an Izmir software studio implements the EU AI Act in product code: feature-level risk tiers, audit logs, human oversight UI, and KVKK overlap.

Jun 17, 2026 Eren Bağcı

Small Language Models We Actually Ship: On-Device AI Without a Cloud Bill

When Gidysoft ships small language models on device instead of cloud LLMs — offline learning apps, cost, latency, and a fallback path that stays honest.