Case Studies

Real work, real outcomes.

A look under the hood at how we approach hard problems — the decisions, the dead ends, and the results that shipped.

AI & Machine LearningJun 2026

Building a Safety Test Harness for a Voice AI Agent

Voice AI agent — a reflective voice companion

An automated voice harness that stress-tests a voice AI agent’s crisis-safety routing on real audio, judges it on the agent’s exact words, and proved it routes self-harm, abuse, psychosis, and medication crises correctly — while surfacing a genuine product gap the team chose to track.

Read the case study →
AI & Machine Learning12–14 Jun 2026

Reaching Parity with Claude Opus 4.8 Using a Local 32B Model

Synthesis AI — biologics CDMO CMC proposal generator

A local OLMo-3.1-32B + LoRA adapter reached parity with Claude Opus 4.8 on the hardest in-domain proposal-authoring workflows — private, fixed-cost, and customer-controlled.

Read the case study →
AI & Machine LearningJun 2026

Making a Frontier Model ~52% Cheaper and ~74% Faster — Without Downgrading It

Document-generation AI application for a regulated life-sciences / CDMO workflow

A long-form document-generation app ran every proposal as one long, serial call to a premium frontier model (Claude Opus 4.8) — accurate, but slow and expensive. Rather than swap the model out for a cheaper one, we kept Opus 4.8 and re-engineered how we call it: splitting each document into independent section groups generated concurrently. On the same frontier model, that cut cost ~52% and end-to-end latency ~74%, with output quality held — verified by an independent judge, not by the team that built it.

Read the case study →

Have a hard problem?

Tell us about your project. We'll scope a plan together.

Start a project →