Intelligence, paintedinto your business.
IKARZA is an AI services company. We design the workflows, systems and strategy that let intelligent software carry the work your team shouldn't have to.
The philosophy
Most businesses don’t need more software. They need a clearer picture. We study how your work flows, then we rebuild it: AI woven into every step, with intent.
What we do
AI Strategy
Know where to start.
We map your business to find where time is lost, where judgment is wasted on the repetitive, and where intelligence would change the outcome. You leave with a prioritised AI roadmap, not a slide deck.
- Opportunity & readiness audit
- AI roadmap and business case
- Build / buy / partner decisions
Workflow Automation
Hand off the busywork.
Agentic workflows that carry real work: quoting, onboarding, reporting, support, operations. Each one designed end to end, integrated with the tools you already use, and measured against hours returned.
- Agentic process automation
- Integrations with your existing stack
- Human-in-the-loop by design
Custom AI Systems
Build what's missing.
When off-the-shelf ends, we build: internal copilots, knowledge systems trained on your documents, customer-facing AI products. Engineered properly, owned by you, refined until it disappears into the way you work.
- Internal copilots & knowledge systems
- Customer-facing AI products
- Evaluation, safety & ongoing tuning
The method
Listen
We sit with your team and watch the work happen. No tooling talk yet, just the honest shape of how things flow, and where they don't.
Design
We draft the future state: which decisions stay human, which steps the AI carries, and what it's worth. You approve the plan before we build anything.
Build
We work in short, visible passes. Working software every week, not a reveal at the end. You see progress early and steer while it's easy to change course.
Refine
Launched is not finished. We measure, tune and train your team until the system fades into the background and the results stay in the foreground.
Every engagement starts as a raw stroke,
and ships as a precise system.
Selected work
Open weights · Hugging Face
Gemma 4 12B Uncensored
127,000 downloads a month.
Google's Gemma 4 12B, decensored with directional ablation and rebuilt for local inference. The ablation touches only the language weights, so the unified multimodal stack and 256K context carry over while capability benchmarks hold. Eight GGUF quantization tiers step from full precision down to a 4.5 GB laptop file, with a multimodal projector for image and audio input and a speculative-decoding drafter that lifts llama.cpp throughput up to 2.2×.
- Refusals cut from 99 to 23 per 100, capability held
- Q2_K to F16 quantization ladder, llama.cpp-ready
- MTP draft model + multimodal projector shipped alongside
Product · macOS
Wandshot
Circle it. Gone.
A menubar-native screenshot studio for the Mac: pixel-exact ScreenCaptureKit capture across Retina and multi-display setups, non-destructive vector annotation, and generative removal behind anything you brush out. The image model regenerates the whole frame, so output is composited back strictly inside the brushed mask, and instructions never reach the model raw: they are sanitised into a guarded template that permits removal, redaction and cleanup, nothing else.
- Swift 6 strict concurrency, actor-isolated pipeline
- Mask-bounded compositing, pixel-exact outside the stroke
- Fastify + Zod API keeps keys and prompts server-side
Product · Web
Uchitil
A tutor that remembers you.
A production AI tutor that teaches Russian through live conversation, every reply streamed and spoken by one of three teaching personas. A Python ingestion pipeline turns scanned textbooks into its knowledge base: per-page OCR detection, chunking, then multilingual embeddings into Pinecone, so a student can ask in English and retrieve from Russian source material. Conversational word frequency computed from the OpenSubtitles corpus is diffed against each learner's exposure log to drive recommendations, and an async memory pipeline feeds their struggles back into every lesson.
- OCR to Pinecone: RAG over scanned textbooks
- Cross-lingual retrieval, English queries over Russian content
- Corpus-driven vocabulary gap analysis per learner
Open source · Evaluation
Crucible
What survives quantization.
An open-source evaluation workbench for self-hosted language models. Capability, refusal behaviour, tool calling, RAG and agentic loops, measured against any OpenAI-compatible server. Every run is provenance-hashed and every suite is versioned, so a claim about a model can be reproduced by anyone who doubts it.
- Seven suites, capability to agentic
- Deterministic grading plus LLM judge
- Delta reports with regression gating
The best place to start
is a conversation.
A 30 minute call costs nothing and usually surfaces two or three workflows worth automating. No pitch, no deck, just a conversation about your business.
Book a call