Antali Labs

Good software, with AI where it helps.

Antali Labs is a small software and AI lab. We build agents and multi-agent workflows, machine learning models, apps, and the cloud setup that keeps them running.

See what we're buildingQuestions? Say hello

Any framework

Pick the framework that fits the job. Underneath, every agent gets the same foundations.

  • Claude Agent SDKAgents that plan, use tools and check their own work.
  • OpenAI Agents SDKHandoffs, guardrails and tracing built in.
  • Google ADKMulti-agent systems on Gemini and Vertex AI.
  • LangChain & LangGraphStateful graphs with retries and human checkpoints.
  • Something elseCrewAI, LlamaIndex, Pydantic AI, or plain code.
The same foundations
  • mcp tools
  • a2a
  • memory
  • evals
  • guardrails
  • traces
  • AWS
  • Google Cloud
  • Your app

And the rest of the toolkit

Claude Agent SDKAgents
PyTorchDeep learning
LangGraphAgent workflows
Next.jsWeb apps
A2AAgent to agent
AWSCloud
Hugging FaceOpen models
PostgreSQL & pgvectorData and vector search
MCPTool connections
LlamaIndexRetrieval
FastAPIAPIs
TypeScriptFront and back end
Google CloudCloud
OpenAI Agents SDKAgents
TensorFlowModel training
Neo4jKnowledge graphs
Google ADK & GeminiAgents
LangChainLLM apps
RedisCaching
CrewAIAgent crews
Scikit-learnClassic ML
n8nAutomation
AWS BedrockHosted models
DockerContainers
SwiftApple apps

Agents working together

Bigger jobs go to a team. An orchestrator splits the work, specialists handle their part, and a person approves what matters.

A2AA2AMCPMCP
Customer
Orchestrator
Billing agent
Policy agent
Payments
Help centre
A person
Event log1/7
  1. customerI was charged twice for order 48213.

Agents talk over A2A

Specialist agents, even ones written with different frameworks, hand each other tasks through the open Agent2Agent protocol.

Tools plug in through MCP

Payments, documents, calendars: each is an MCP server. Adding one doesn't mean rebuilding the agents.

A person signs off

Anything that moves money or can't be undone waits for approval, and every step lands in the log.

What we build

From a single agent to the app around it and the cloud it runs on. Pick one to see it work.

01orchestrator → 3 specialists over A2A
research
billing
support
  • a2a → research find the order history

Answers from your documents

Agents answer from what you've written, not from memory. A question is matched by meaning, and every answer shows where it came from.

›

embedsearchrerankanswer
Embedding space · 1,284 chunkscosine

The question becomes a vector and lands next to the chunks that mean the same thing.

Top 3 after rerank

    Search by meaning

    Documents are split into chunks and turned into embeddings, stored in pgvector or a managed vector index. A question finds the chunks closest to it in meaning, not just the ones that share its words.

    Hybrid, then reranked

    Vector search runs alongside keyword search, and a reranker puts the best passages first before the model sees them.

    Every answer is checkable

    Each sentence points back to the chunk it came from, and retrieval has its own evals, so a bad match shows up before a customer does.

    An agent at work

    An agent is an AI model that can look things up and take action. Pick a job and watch one work.

    Example

    My order still hasn't arrived. Can you check?

    Tools it used

    1. Found the orderfind_order

    2. Checked the trackingtrack_parcel

    3. Booked a redeliverybook_redelivery

    From prompt to production

    The hard part is making an agent dependable. So every one goes through the same loop: build, test against real cases, ship, and watch it.

    Traces become new eval cases. Back to step 2.

    support-agent/agent.pyBuild

    Currently

    Most of what we build runs quietly behind someone else's product. This one's ours.

    iOS appIn development

    ReForm

    An AI coach for eating well and training, made for iPhone.

    Platform
    iPhone, iOS 26
    Built with
    SwiftUI, Claude SDK, Gemini API
    Status
    Not out yet

    More in the lab. A couple of other apps are on the bench. They'll show up here once they're ready.

    How we work

    1. Keep it simple

      We pick tools that are well understood and skip the layers nobody needs. Less to break, less to maintain.

    2. Make AI predictable

      Agents get clear rules, checked outputs, evals that have to pass, and a way to recover from mistakes.

    3. Data stays where it belongs

      Everything can run inside one cloud account. Records and models don't have to leave it.

    Have any doubts?

    If something here raised a question, or you just want to talk shop, this inbox reaches us directly.

    contact@antalilabs.com
    Say hello