The stack

The tech stack we build, run and trust in production.

Model-agnostic. Tool-rich. Observed end-to-end. We compose the right layer per workflow, and bring guardrails, evals and human oversight from day one.

Models

Frontier and open models, picked per task to balance quality, latency, cost and safety.

OpenAI GPT-5 / o-series
Anthropic Claude Sonnet & Opus
Google Gemini 2.x
Meta Llama, Mistral, Qwen (self-hosted)
Whisper, Deepgram, ElevenLabs (speech)

Agent frameworks

Graph-based orchestration with explicit state, tools, retries and human checkpoints.

LangGraph & LangChain
Mastra, CrewAI, AutoGen
OpenAI Agents SDK
Vercel AI SDK
Custom orchestrators when frameworks get in the way

Tools & integrations

Typed, permissioned tools that let agents act, not just answer.

MCP servers (custom + open source)
Function calling / structured outputs
EHR & FHIR connectors
Twilio, Slack, Gmail, Calendar, Stripe
Browser automation (Playwright, Browserbase)

Retrieval & memory

Grounded answers with hybrid retrieval, citations and per-document access control.

pgvector, Pinecone, Weaviate, Qdrant
Hybrid BM25 + dense retrieval
Reranking (Cohere, Voyage)
Long-term memory stores (Mem0, Letta)
Document parsing: Unstructured, LlamaParse

Realtime voice

Sub-second voice agents with barge-in, function calling and human warm-transfer.

OpenAI Realtime, Gemini Live
LiveKit Agents, Pipecat, Vapi
Twilio Voice, Telnyx
Deepgram & Whisper STT
ElevenLabs & Cartesia TTS

Evals & observability

Continuous evaluation and tracing ensure agents are tested like systems, not vibes.

LangSmith, Langfuse, Braintrust
Helicone, Arize, OpenTelemetry
Golden datasets + LLM-as-judge
Regression and drift alerting
Per-trace cost & latency budgets

Security & compliance

BAA-covered providers, PHI minimization and tenant isolation by default.

BAAs with OpenAI, Anthropic, AWS, GCP
PHI redaction & data minimization
RBAC, audit logs, tenant isolation
SSO / SCIM, KMS-managed secrets
HIPAA + SOC2-aligned controls

Workflow & data

Pipelines, queues and warehouses that feed agents and absorb their outputs.

Temporal, Inngest, Trigger.dev
Postgres, Supabase, Redis
BigQuery, Snowflake, dbt
Kafka & event streams
Airbyte / Fivetran connectors

Infra & deployment

Edge-fast, GPU-ready and cost-tuned to run wherever your data is allowed to live.

AWS (Bedrock, ECS, Lambda)
GCP Vertex AI, Azure OpenAI
Cloudflare Workers / Vercel edge
Modal, Replicate, RunPod for GPUs
Docker, Terraform, GitHub Actions

Foundations

The engineering and ML stack that hosts the agents.

Agents don't live in a vacuum. They sit inside real apps, services and data platforms, and not every workload needs an LLM. Here's the rest of what we build and run on.

See agent services

Product engineering