AI agents that make it to production.
Humanitas Studio is a San Francisco AI engineering studio. We design, build, and operate LLM-powered agents and retrieval systems for businesses, and we stay accountable for them after launch.
Two agents in production for a San Francisco signage firm
Our client is a San Francisco architectural signage firm with decades of projects behind it, and decades of institutional knowledge locked in its website, catalogs, and photo archive.
The foundation: a living knowledge base
With the client’s authorization we ingested the company’s full web presence and ran AI vision analysis across its sign portfolio. The result is a vector knowledge base that refreshes itself every day, rebuilds without ever destroying good data, and rolls back on its own if a refresh looks wrong.
Agent one: the internal virtual consultant
A private assistant the client’s sales and marketing teams use every day: semantic search with intelligent filtering, portfolio statistics on demand via SQL tool use, and multi-turn conversations grounded strictly in the client’s own corpus.
Agent two: the public concierge
A named, friendly assistant that answers visitors’ questions directly on the client’s website, with a persona and content boundaries approved by its CEO. It replaced a generic third-party widget in June 2026 and has represented the brand ever since.
Every model choice is a balance of four forces.
New models arrive every month, and the loudest leaderboard rarely reflects your workload. Before a model earns a place in your system, we run the candidates head-to-head on your real tasks and weigh four things against each other:
Intelligence
Does the model actually get your domain right? We measure answer quality on your own data and your own questions, not on public benchmark scores.
Speed
A customer-facing assistant lives or dies on response time. We measure real latency at every step and engineer it down where the model alone falls short.
Cost
What matters is the cost of a conversation at your real volume, not the price per token. We project a month of actual usage before you commit to anything.
Privacy
Which provider sees your data, and under what terms. Some information should never leave your systems at all, and the architecture has to respect that.
The right balance is different for every business, and often for every feature inside the same product. The two agents we run for this client sit on different models for exactly this reason: each one earned its place in its own bake-off. Claude is our home base, and even Claude has to keep winning its seat.
The whole system, from corpus to conversation.
Agentic systems
Multi-turn assistants that use real tools: semantic search, SQL analytics, structured workflows. We build primarily on Anthropic’s Claude, with prompt caching to keep them fast and affordable.
Retrieval & knowledge bases
We turn scattered websites, catalogs, and image archives into curated vector knowledge bases that refresh themselves. If your knowledge lives in photos, we run AI vision analysis over those too.
Customer-facing assistants
Public chat concierges with a brand-approved persona and strict content boundaries, embedded directly in your site. Far more capable than an off-the-shelf widget.
Model strategy
Provider-agnostic architecture. Every model has to earn its seat in a bake-off, judged on intelligence, speed, cost, and privacy.
Technology in service of people.
Humanitas is the Latin word for human culture and kindness. We chose the name because it says what the work is for. A good AI system makes the people around it more capable, and the judgment that makes your business what it is stays yours.
Humanitas Studio is led by Christian Sasso, founder and principal engineer. We keep the studio deliberately small: the person who designs your system is the person who builds it, ships it, and answers for it.