AI / ML research, built to ship.

We design, evaluate, and deploy machine-learning systems — from agentic LLM applications to retrieval pipelines and the MLOps that keeps them honest in production.

Start a project → Security bay

Experiment

Training monitor EXP-0143

Val lossep 12

0.061

Throughputtok/s

21.4k

GPU util%

94avg

Eval accuracyheld-out

0.941

FEED: SIMULATEDREFRESH 1.0s

01 Focus areas

From frontier idea to running system

We don't stop at a demo. We measure, harden, and deploy — then keep iterating against real-world signal.

Agentic systems — tool-using LLM applications with guardrails
Retrieval pipelines — RAG, knowledge graphs, hybrid search
Fine-tuning & eval — adapters, harnesses, red-teaming
MLOps — deployment, observability, cost control
Vision & structured data — CV and tabular modelling

eval.py — vair-ai

# benchmark a candidate pipeline
from vair import Pipeline, Eval

pipe = Pipeline("rag-v3")
report = Eval(pipe).run(
suite="production",
redteam=True)

# → accuracy 0.94
# → p95 latency 380ms
# → safety PASS ✓

02 Bench

Instruments in this bay

Tensor activationslayer 12 · head 3

Context memory0x4000

Build

Greenfield

AI features and products designed and delivered end-to-end.

Evaluate

Independent

Benchmarking and red-teaming of your existing models and agents.

Operate

In production

MLOps, monitoring, and continuous improvement once you're live.

Stack: PythonPyTorchClaude / LLM APIsVector DBsDockerEval harnesses

03 Lab log

Bay 01 experiments

2026-09-11 08:14ZEXP-0143Fine-tune run — domain adapterEpoch 12/20 · val loss 0.061 · no driftRunning

2026-09-10 21:02ZEXP-0142RAG-v3 evaluation suiteAccuracy 0.94 · p95 380ms · safety passPass

2026-09-08 09:15ZEXP-0141Agent harness v2.3 regression sweep412 tasks · 0 regressionsReview

Open channel

Have a model or agent to build?

From prototype to production — let's scope it.

Talk to the AI team →