From the engineering team
Practical insights on building AI systems that scale — no filler, no marketing fluff.
Agent Design Patterns: A Practical Guide to Building Reliable AI Agents
A working vocabulary of reasoning and orchestration patterns for building AI agents that don't fail in production.
Postgres as your AI memory layer: pgvector in production
pgvector is good enough for most AI memory workloads if you tune the indexes, partition the schema, and stop pretending it's a dedicated vector DB.
LLM function calling patterns that survive real users
Function calling breaks in production not from bad models but from bad schemas, missing idempotency, and trusting LLM-supplied arguments — here's what holds up.
SOC 2 for AI startups: what actually matters in year one
Most AI startups over-invest in SOC 2 controls auditors don't care about and under-invest in the evidence pipeline that actually fails the audit.
Why we use Temporal for every long-running AI workflow
Long-running AI workflows fail in ways HTTP retries can't fix; Temporal's durable execution model is the only thing we've found that survives production.
Controlling LLM costs in production before they control you
LLM costs scale with usage, not revenue — control token waste, caching, model routing, and observability before the bill outpaces your margins.
Multi-tenancy patterns that don't blow up at scale
Pool, silo, and bridge isolation each fail differently at scale — pick based on blast radius, noisy-neighbor tolerance, and per-tenant cost.
LangGraph in production: what the docs don't tell you
LangGraph's tutorials get you a demo; production exposes state bloat, checkpoint contention, and silent retry loops the docs never mention.
How prompt caching cuts your LLM bill by 60%
Prompt caching reuses tokenized prefixes across requests, cutting input costs 50-90% on workloads with stable system prompts and shared context.
The 200ms voice latency budget — where every millisecond goes
A frame-by-frame breakdown of a sub-second voice agent's turn budget. STT, network, LLM, TTS — and the surprising places we've shaved 80ms.
Your eval set is your spec — write it before the prompt
Why the most expensive AI mistake we see at customer engagements is teams tuning prompts before they've written the regression test that defines 'right'.
How we design multi-agent systems for production, not demos
The orchestration patterns that survive contact with real customers — and the demo-ware that doesn't. Drawing on a year of LangGraph in production.
Where RAG stops being RAG and starts being a search problem
After two dozen RAG deployments we've stopped calling it RAG. Here's the search and retrieval stack that actually works in production.
Shipping a real MVP in 14 calendar days — and why most teams can't
The pre-engineered scaffolding that makes a 14-day MVP possible. Auth, billing, RBAC, audit, deploy — already done before kickoff.
The case for boring infrastructure under interesting AI
Postgres, Redis, Temporal, Terraform. Why we pick technology that will be running in five years over the framework trending this quarter.
Voice AI compliance: what HIPAA actually requires (and doesn't)
A field guide to the voice AI compliance questions we hear most often from healthcare CIOs — and the misconceptions to leave behind.
Operating agentic systems: the on-call surface no one warned you about
Agents introduce a new category of incidents — drift, runaway loops, tool-use failures. Here's the runbook we've evolved over a year.
Vendor or build: an honest decision tree for AI features
When to buy OpenAI's stack as-is, when to wrap, when to fork. The framework we use with CTOs every week.
Ready when you are
Want this thinking applied to your stack?
Book a call and bring your thorniest architecture question. We'll give you an honest take — on us.