Applied AI engineering studio
We design and deliver fine-tuned models, retrieval pipelines, agentic workflows and computer vision systems for B2B teams that need outcomes, not slideware.
What we build
We pick the smallest set of tools that solves your problem. No buzzword-driven architecture.
Adapt open-weight models to your domain, data and terms of use. LoRA / QLoRA for fast iteration, full fine-tunes when warranted, preference tuning (DPO) where the data supports it.
Turn internal knowledge into a queryable system. Hybrid retrieval, evaluation pipelines, observability, and MCP servers your agents and IDEs can call directly.
Coordinated agents that plan, act, recover and stay in scope. Tool use, memory, guardrails and human-in-the-loop hand-offs — production, not demos.
Adversarial review of your stack — web, network and AI-specific. Prompt injection, jailbreak, model and prompt extraction, supply-chain risks across the LLM lifecycle.
Custom models for measurement, detection and broadcast analytics. From edge devices to broadcast-grade pipelines, designed for the metrics that matter to your operation.
How we work
A predictable engagement model that lets your team see traction in weeks, not quarters.
01 · Discovery
A 60–90 minute working call. We map the outcome, constraints, available data and what "shipped" looks like for your team.
02 · Pilot
A scoped 2–6 week pilot that proves the approach on your data. Fixed scope, fixed timeline, written hypothesis.
03 · Delivery
Ship into your cloud or on-prem with your guardrails. Documented, observable, owned by your team — not by us.
04 · Support
Optional retainer for updates, eval suites, and keeping pace with the rate at which models — and your data — drift.
Selected work
Public case studies are in progress. Below is a short selection — happy to discuss specifics under NDA.
Per-frame player and ball tracking, event detection and broadcast overlays running at production latency for live sports.
End-to-end control software for mobile parcel lockers — device firmware bridge, operator dashboard, customer flow and analytics.
Real-time speech-to-speech translation from video, preserving speaker characteristics. Built for long-form content and live streams.
About
DeQuzzy was founded on 2 November 2023 in Nha Trang, Vietnam. We are a deliberately small team of senior engineers who have shipped applied AI, classical computer vision and secure systems for B2B clients across Asia and Europe.
We pick projects where modern AI is the unfair advantage — not the marketing layer. Vietnamese, English and Russian working languages. Remote-first, but happy to fly to you when it matters.
Contact
The form is being rebuilt. In the meantime, the fastest channel is e-mail — or open the channels below.