What we build
We build RAG systems, agent workflows, evaluation pipelines, ETL, and ML serving. Most projects ship behind a feature flag with an offline eval suite. We're cynical about magic and serious about cost.
RAG systems
Retrieval, chunking, reranking, and eval — not a vector DB lobby.
Agent workflows
Tool-using agents with observability, retries, and human-in-the-loop.
Data platforms
Warehouse, lake, dbt models, BI dashboards.
Model serving
Self-hosted Llama / Mixtral with vLLM behind a managed API.
Evaluation harness
Offline + online eval so you can ship a prompt change without crossed fingers.
The tools we'll reach for.
- FastAPI
- LangChain
- vLLM
- Pinecone / pgvector
- Python
- dbt
- Postgres
- BigQuery
- Airflow
- GCP
- AWS Bedrock
- Modal
- Sentry
How we work in this area
Before we touch a model, we write the 50-prompt eval that says 'shipped'.
RAG / agent / fine-tune — whichever survives the eval.
Cost dashboards + drift alerts on day one.
Selected work in this space
FAQ
GPT-4 or open source?+
Closed for v1; open for cost or compliance once the eval is steady.
Do you fine-tune?+
Rarely. RAG + good prompts win 80% of the time.
Can you build agents?+
Yes, with tool-use, retries, and a human-in-the-loop seat.
How do you keep costs sane?+
Caching, smaller models for routing, and dashboards. We track $/request like latency.
Haveaai&dataproject?
Tell us where you are and we'll tell you what's possible in the first two weeks.
next kickoff · week 34