TBC-M8
Enterprise workhorse
The backbone of multi-agent scenarios. Long context, reliable tool calling and structured output generation carry production workloads.
A TBC TEKNOLOJI COMPANY · SINCE 2019
Your own model.Your ownagents.On your own infrastructure.
TBC Tech AI builds SLM/LLM models trained on your organisation's data, produces the AI agents that do real work on top of them, and runs them all from a single control tower — in the cloud, on your servers, or fully offline.
- Agent types in production
- 24+
- Enterprise model families
- 6
- Orchestration availability
- 99.9%
- Data sovereignty option
- 100%
LIVE ORCHESTRATION
What happens when a task arrives? Router, planner, tool agents and critic engage within seconds.
The flow below represents how the TBC Agent Runtime actually works — it runs while you watch.
tbc-agent-runtime · live
- Input / Output
- Core
- Task agent
- Control
trace · --------0 ms
- Tokens
- 0
- Cost
- $0.000
- Tool calls
- 0
- Guardrail
- OK
PLATFORM LAYERS
One stack: model, memory, tools, runtime, evaluation and observability. Touch a layer to open it.
Click a layer
MODEL FAMILY
A bigger model does not solve every problem. We run the right-sized model, on the right data, in the right place.
With an SLM (Small Language Model) approach: lower cost, lower latency, and data that never leaves your perimeter.
8B
TBC-S1
Edge / routing
The smallest member, running even on CPUs. It finishes high-volume work — classification, routing, summarisation, masking — in under a second.
TBC-S3
General purpose
An in-house assistant on a single GPU. A balanced option for document Q&A, drafting and tool calling.
TBC-R1
Reasoning
For tasks that need thinking time: multi-step planning, root-cause analysis and complex reconciliation.
TBC-MED
Healthcare domain
A variant adapted to clinical text, medical terminology and report structure. Runs without patient data leaving the institution.
TBC-FIN
Finance domain
Trained on invoices, statements, reconciliation and regulatory text, with a focus on numerical consistency.
Scores show relative performance on our internal comparison sets; they are re-measured on your own data during the project.
MODEL PICKER
Pick your scenario — we will size the setup instantly.
RECOMMENDED SETUP
TBC-S3
An in-house assistant on a single GPU. A balanced option for document Q&A, drafting and tool calling.
- Primary model
- TBC-S3 · 3.8B
- Router
- TBC-S1 1.5B
- Hardware
- 1 × GPU 24 GB
- Expected response
- ~ 0.8 s
- Capacity
- ~ 0.7 req/s
- Time to deploy
- ~ 3 weeks
This is a preliminary estimate based on average prompt length, RAG usage and typical hardware efficiency. Exact sizing comes out of the discovery session.
AGENT FACTORY
We move an agent from idea to production in six steps — each with a measurable output.
Discovery
We observe the process on site. Any step that is repetitive, rule-bound and data-fed is an agent candidate.
Scope and success criteria
Data Foundation
Documents, databases and process records are collected, cleaned, labelled and made retrievable.
Corporate knowledge base
Model
The right-sized model is chosen, adapted to your language and fine-tuned on domain data where needed.
Domain-adapted model
Agent Build
Tools are wired, steps are defined, failure and approval paths are written. The agent talks to real systems.
A working agent
Evaluation
An exam against the golden set. Accuracy, source faithfulness, latency and cost are measured together.
Go-live decision report
Production & Operations
Staged rollout, monitoring, cost control and continuous improvement. Not a handover — joint operation.
Live system with an SLA
READY AGENTS
Do not start from zero. We adapt agent templates that already work in your sector to your own data.
Document & Contract Agent
General
Reads contracts, proposals and correspondence; extracts risky clauses, dates and amounts, compares versions and summarises.
- 12 s/document
- 96% field extraction
- OCR
- RAG
- DMS
- e-Imza
Customer Support Agent
General
Answers calls, e-mails and chats from one knowledge base, and hands off what it cannot solve with a summary for the right team.
- 68% first-contact resolution
- 24/7
- CRM
- Ticket
- Whisper
- TTS
Code & Test Agent
General
Analyses legacy systems, writes tests, reviews code and drafts modernisation steps.
- 40% faster review
- Git
- CI
- Static Analysis
Data Protection Agent
General
Scans personal data across systems, keeps the processing inventory current and reports risky areas.
- Automated inventory
- Monthly report
- PII Scan
- Policy
- Audit
Tender & Bid Agent
Public sector
Monitors notices, reads specifications, assesses your eligibility and drafts the skeleton of the bid file.
- Daily scan
- Eligibility score
- EKAP
- Scraper
- Alert
Clinical Report Agent
Healthcare
Structures patient notes and test results, suggests coding and reminds the clinician of missing fields.
- ICD-10 suggestions
- On-prem
- HL7 FHIR
- ICD-10
- PACS
Laboratory (LIS) Agent
Healthcare
Catches outliers in results, suggests retests and raises critical values immediately.
- Instant critical alert
- LIS
- Anomaly
- Report
Financial Reconciliation Agent
Finance
Matches e-invoices, bank statements and ERP records; finds the difference and writes the reason.
- 94% auto-match
- End of day
- e-Fatura
- ERP
- SQL
Credit & Risk Agent
Finance
Compiles the application file, reads financial statements, scores it and explains the decision rationale.
- Explainable score
- Scorecard
- KKB
- XGBoost
Shipment Planning Agent
Logistics
Aggregates orders, optimises routes and vehicle capacity, and flags delay risk in advance.
- 17% km saved
- Route API
- GIS
- Forecast
Predictive Maintenance Agent
Manufacturing
Tracks trends in sensor data, warns before failure and opens the maintenance work order automatically.
- 48 h before failure
- SCADA
- IoT
- Time-series
Visual Quality Agent
Manufacturing
Inspects products on the line with cameras, classifies defects and writes back to the production system.
- 98.6% detection
- Runs at the edge
- YOLO
- Edge GPU
- MES
SECTOR SOLUTIONS
TBC Teknoloji's enterprise software experience, combined with an AI layer.
Healthcare
We combine our experience in LIS and PACS integrations with local models that keep patient data inside the institution — shifting clinicians' time from documentation back to patients.
- Less reporting time
- 38%
- Faster coding
- 6x
- Patient data leaving
- 0
- 🩺Clinical documentationTurns dictated notes into structured reports.
- 🩺Test triageMoves critical results to the front of the queue.
- 🩺Appointments and callsHandles patient routing 24/7.
- 🩺Regulatory complianceGuarantees data stays on premises.
DEPLOYMENT MODELS
Wherever your data must live, the model runs there.
TBC Cloud
Fastest start
A tenant-isolated environment in our data centre in Türkiye. We manage the hardware, you focus on the work.
- Live within days
- Autoscaling and redundancy
- Usage-based cost
- Managed updates and monitoring
Your Own Servers
Data stays inside
We install the models in your data centre. No request leaves the corporate network and the hardware stays yours.
- Data never crosses your perimeter
- Fits your existing virtualisation
- Fixed, predictable cost
- Corporate SSO and directory integration
Air-Gapped
No internet required
Full functionality in disconnected environments for public sector, defence and critical infrastructure. Updates travel on controlled media.
- No external network needed
- Controlled update packages
- Local model and vector store
- Fits physical access controls
SECURITY & COMPLIANCE
AI cannot sit outside enterprise risk management.
Every request is identified, every output reviewed, every decision traced. Personal data is masked before it ever reaches the model.
- ✓KVKK-compliant process
- ✓GDPR-compliant by design
- ✓ISO 27001-aligned operations
- ✓Data centre in Türkiye
- ✓Open-source based, vendor independent
- 🥷
PII Masking
Identifiers, IBANs, phone numbers and names are detected and masked before entering the model prompt.
- 🔑
Role-Based Access
The agent runs with the user's permissions. What the user cannot see, the agent cannot reach.
- 📜
Audit Trail
Which prompt, which source, which tool and which output — kept in immutable records.
- ⚖️
Policy Engine
Corporate rules are enforced at output level; a response that violates policy never reaches the user.
- 🔐
End-to-End Encryption
TLS in transit, disk-level encryption at rest and separated key management.
- 🧱
Tenant Isolation
Each organisation's model, vector store and logs run separately; data never mixes.
WHY TBC TECH AI
Plenty of people can demo. Few take it to production, operate it and own the outcome.
Engineering that goes on site
Research only counts once it ships. Our team observes the process in your office, builds the agent on your real data, and stays on site until it is live.
One stack, end to end
Model training, agent runtime, integration, security and monitoring from the same team. No stitching five vendors together.
-64%
Right size, lower cost
We do not send every task to a giant model. Work a small model can do stays on the small model.
3 hf
First agent in three weeks
Our typical time from discovery to a measurable pilot output.
100%
Data sovereignty
If you want it, not a single bit leaves your organisation.
An enterprise software background
As TBC Teknoloji we have been writing critical systems for healthcare, finance, public sector and defence since 2019. We build AI into those systems, not on top of them.
Joint operation
We do not deliver and disappear. SLA, on-call and continuous improvement included.
FAQ
The questions technical teams ask most.
Is our data shared with others during model training?
No. A model trained on your data belongs only to you; it is not transferred to another customer's model, a shared pool or a third-party provider. In on-prem and air-gapped deployments the data never leaves your network at all.
Why a small model (SLM) instead of a large off-the-shelf one?
Most enterprise work is narrow and repetitive: document classification, field extraction, routing, summarisation. On those tasks a domain-adapted 3–8 billion parameter model approaches or beats a giant model at far lower cost and latency. For steps that genuinely require complex reasoning we bring in the large model — the choice follows the work.
How do you connect to our existing systems (ERP, LIS, DMS)?
Agents act through tools. Every integration is modelled as a tool with a defined schema, authorisation and logging. We use REST, SOAP, database views, file shares and — where needed — an RPA bridge. For legacy systems without an API, TBC Teknoloji's modernisation experience steps in.
What happens when the model answers incorrectly?
Three layers apply: the agent cannot answer without citing a source, a critic compares the answer against those sources, and the guardrail layer stops output that violates policy. For critical processes we also define a human approval step — the agent prepares, an authorised person approves.
How long until we go live?
For a typical first agent, three weeks from discovery to a measurable pilot. The number of integrations, data quality and approval cycles affect this. We recommend keeping the first scope small and measurable.
What hardware is required?
1.5B models can run on a modern server CPU. For the 8B class a single 24–48 GB GPU is typically enough; we scale horizontally under high concurrency. The model picker above gives a preliminary estimate for your scenario.
Can our own team take the system over?
Yes. We use an open-source based stack; model weights, prompts, evaluation sets and infrastructure code are yours. Handover training and documentation are part of delivery.
What is the relationship between TBC Tech AI and TBC Teknoloji?
TBC Tech AI is the AI-focused company of TBC Teknoloji, which has been building enterprise software since 2019. That experience in healthcare, finance, public sector and defence is the foundation of our AI products.
CONTACT
A 30-minute technical discovery call.
We listen to your processes, work out together which step can be automated with an agent, and define the pilot scope. Not a sales pitch — an engineering conversation.