Ran the KANGAL autonomy competition from start to finish.
Registrations, juries, the scenario engine and every schedule change for an unmanned ground vehicle and simulation contest.
We build the language models, GPU infrastructure and AI platforms that defense and public institutions run inside their own walls.
From the GPU racks to the assistant your analysts talk to, we design, build and operate every layer. The model, the data and the answers all stay on your network.

Ask in plain language and get answers grounded in your own documents, with the source passage attached. Semantic search, summaries, report drafting and document classification in one interface.

Each request goes to the model that fits the task and the data, with load balancing, priorities, RAG and MCP.

Our own Turkish MoE language model, Resif, plus pre-training, fine-tuning and distillation with Turkish-first tokenizers.

GPU clusters run as one HPC system: one Slurm queue, NVLink inside nodes and InfiniBand between them, no Kubernetes needed, and models moved into closed networks by controlled transfer.
A workspace for every model, agents that do the work, a council that decides, an assistant on the desk and an evidence engine for open sources. All written by our own team, all able to run without internet.

One gateway in front of every model, with assistants, knowledge bases, workflows and MCP tools built on top of it.
Explore AIFab
Twenty Turkish-first agents that plan, remember, use your tools and work on a schedule.
Explore AgenFab
A council of AI expert personas that research with open sources and your documents, debate, and hand you a verdict.
Explore Consensus
The agent platform on one desk: cited answers from your documents, Office and Outlook work, and MCP links to your applications, on a closed network.
Explore Dayı
Open-source research that becomes evidence: isolated collection, archive first, and every claim tied to a verbatim quote.
Explore SearchFabDeveloped for the Presidency of Defense Industries (SSB) and run together with it, on open-source foundations that add no licence dependency.

| Model | Status | GPU | Progress |
|---|---|---|---|
| vehicle-det-v7 | Running | 8× | |
| doc-classifier-tr | Running | 4× | |
| thermal-seg-v2 | Queued | 2× | |
| plate-ocr-v4 | Published | 1× | |
| drone-track-v3 | Published | 4× |






Behind every model is the vision, software, electronics and site work that keeps it running at military sites, in steel plants, in mines and along rail lines.
See projects
Language models on closed networks, RAG and MCP expert systems, institutional assistants.

Real-time detection of people, vehicles, faces, plates and behaviour for borders, critical sites and cities.

Microservice systems for thousands of concurrent users, signed in through LDAP, SSO or e-Devlet.

Multi-axis motion platforms and VR training rigs, driven by control electronics we design.

Custom field devices over Modbus, RS485 and LoRaWAN, with thousands of sensors managed from one map.
Contracted work for defense, public institutions and heavy industry, delivered from architecture to operation.
Registrations, juries, the scenario engine and every schedule change for an unmanned ground vehicle and simulation contest.
Scenario writing, multi-stage juries with weighted scoring for experts from SSB, TSK and ASELSAN, and the award ceremony.
Portals, the shared GPU pool, installation and maintenance for the Defense Industry AI Platform.
Microservices with live scoreboards, accuracy analysis and speed measurement under heavy load.
Climate, power, vibration, smoke, gas, flood and door sensors over Modbus, with alarms that trigger on their own.
3-DOF crane and forklift platforms whose joystick and steering-wheel controls swap in minutes.
Dual-joystick consoles, emergency stops and vibration feedback, ready for VR training software.
Our own boards and sensors, 22-inch panel PCs and predictive-maintenance data from machines that had none.
Billing, subscriptions and SCADA reservoir levels in a single BI layer for senior management.
T.C. SSB
SimBT9-DOFWireless weapon tracking for VRDatasets, benchmarks and quantised models on Hugging Face, so open models work better in Turkish. Our Turkish MoE language model, Resif, and its tokenizer are next.

How much more GPU time the same content costs in Turkish than in English on open models served with vLLM, how many tokens seven tokenizers spend on it, and what FP8 and GGUF quantisation do to Turkish.
Decisions on customer messages, invoices and logs that turn on negation, the -DI/-mIş suffixes, Turkish number formats and official holidays, with and without planted instructions.
How often coding agents submit without running a test or script after their last edit, and how much less often those submissions solve the issue.
GGUF files of Turkish models, quantised with an importance matrix computed on Turkish text, with the loss measured for every file.
Follow the work on Hugging Face.
Every system is designed to work with no outside connection, offline and end-to-end encrypted, to defense and public-sector security standards.
Open-source foundations and database-independent designs. What we deliver is yours to keep.
Boards, firmware, models and platforms come from our own team, not from a reseller's catalogue.
LDAP, SSO and e-Devlet sign-in, API-first services, CI/CD and observability from day one.
CF-MT24W24-DOF wireless motion tracking.
CF-SPD64Controls 6-DOF motion platforms.
CF-F7 AIOAll-in-one flight control for multi-rotor aircraft.
CF-LHGOpen-source outdoor gateway over Wi-Fi, Ethernet or LTE.