AI GAMES DPK · SOFIA, BULGARIA
We design, build and ship AI-powered software end-to-end — web, mobile, data, security and the cloud or on-prem infrastructure beneath them. One team, one accountable contract, no agencies between you and the people who wrote the code.
A small studio with a deliberately wide surface area. Eight practice areas, one accountable team, one signature on the work. We standardise on Cloudflare for our own products and pick the cloud (or on-prem footprint) the project demands for client engagements.
LLM orchestration and ensembles, retrieval-augmented generation, LoRA / QLoRA fine-tuning of open-weight models, calibrated probabilistic forecasting, identity-preserving image generation, vision-language-action models, prompt-injection defences. AI runs on queues, not in request handlers.
Pipelines on Databricks, Snowflake, Azure Synapse and AWS Lake Formation. Data lakes, warehouses and lakehouses; streaming ingestion; embedding pipelines for RAG; LLM-assisted analytics that let business users ask questions of warehouse data in natural language.
Cloudflare, AWS, Azure and GCP solution architecture, plus hybrid and on-prem designs. Multi-cloud reference architectures, lift-and-shift and re-platform migrations, well-architected reviews, FinOps and cost engineering. Infrastructure as code with Terraform, Pulumi or platform-native tooling.
Server, network and storage design — including SAN and NAS bring-up, virtualization stacks (VMware, Proxmox, KVM), capacity planning, backup and disaster recovery. Useful when latency, sovereignty or regulation makes the cloud the wrong answer.
CI/CD that ships every commit, GitOps deployment patterns, infrastructure as code, secrets management, SBOM generation, supply-chain scanning, observability with OpenTelemetry, and policy-as-code gates that fail builds rather than write reports nobody reads.
Threat modelling, SAST and DAST, dependency and container scanning, GDPR and ISO-27001 alignment, secure SDLC reviews. AI-assisted vulnerability triage and auto-remediation: LLM-augmented pipelines that open patched pull requests rather than ticket queues nobody resolves.
Next.js / React on the edge or any cloud, native iOS in Swift / SwiftUI, native Android in Kotlin / Jetpack Compose. PKCE OAuth, generated OpenAPI clients shared across platforms, accessibility (WCAG 2.1 AA) by default — not as a retrofit.
ROS 2 and ArduPilot integration, Bayesian probabilistic decision engines, multi-agent coordination, on-robot VLA manipulation policies (SPEAR-1 by INSAIT on Franka and WidowX). Vendor-agnostic software layers over mixed civilian fleets — search and rescue, agriculture, logistics, environmental monitoring.
A few in-house products in active development — public proof that the studio ships the kind of work we sell. The three below are not all of it; client engagements and other in-house projects are not shown.
Consumer AI — portrait stylizer for web, iOS and Android.
Identity-preserving style transfer for rock-culture portraits. Queue-based generation, payments, mobile apps — built end-to-end in the studio.
Next.js · Cloudflare · FLUX.2
B2B AI — calibrated probabilistic forecasting SaaS.
Multi-model LLM ensemble that produces calibrated probabilities with 80 % / 95 % confidence intervals and a polished PDF report.
Next.js · Cloudflare · OpenRouter
Autonomous systems — coordination layer for mixed civilian fleets.
Vendor-agnostic decision and orchestration platform for drones, robots and unmanned vehicles. ROS 2 + ArduPilot, Bayesian risk scoring.
ROS 2 · ArduPilot · SPEAR-1 · Bayesian AI
Cloudflare for compute, storage, queues, DNS, edge cache and the AI gateway. One dashboard, one CLI, one bill. The studio's leverage comes from depth in one platform, not breadth across ten.
A 90-second model call cannot live inside a 30-second HTTP handler. AI runs on queues; the client gets a 202 in under 100 ms and the work reconciles in the background.
Clients are presentation. Pricing, money, ownership and authority live in code we control — and only there. A tampered phone or forged cookie cannot move funds or skip a check.
Append-only ledgers, idempotency keys, schema-level constraints, generated API clients, lint rules. The system refuses bad states by construction; we never have to remember to do the right thing.
A user in São Paulo or Singapore hits compute next door, not a us-east-1 round trip. Perceived speed is a first-class product requirement, not a late-stage optimisation.
A feature is done when a user can complete it end to end — not when the code merges. Half-built work behind flags is debt. Every sprint ends with one fewer flag, not one more.
AI Games DPK is registered in Sofia as a variable capital company — a legal form designed for startups. That lets us grow at our own pace, without outside investors and without answering to a board.
The studio is intentionally small and plans to stay that way. The same people design, develop and maintain everything we ship — for ourselves and for the clients we choose to work with.
Projects, partnerships, bug reports — one inbox, read by the team.
