Now booking Q3 2026 4 seats left

We ship the AI your team has been promising for two quarters.

A senior studio of ex-OpenAI, ex-Anthropic engineers and applied researchers. We take your AI roadmap from “interesting Slack thread” to a measurable production line in 6–12 weeks.

Book a 30-min discovery cat ./case_studies/
Founded
2022
Models shipped
48+
Enterprise clients
23
Avg. ROI lift
4.1×

In production at

01 — Capabilities

Six disciplines. One studio.

We don’t outsource. Every engagement is staffed by the same team — researcher, applied engineer, designer, ops — across the full delivery.

01

Retrieval & RAG

Embedding, chunking strategy, evals. We build retrieval pipelines that survive contact with the actual corpus — not synthetic benchmarks.

  • OpenAI
  • Anthropic
  • Voyage
  • pgvector
  • Qdrant

02

Fine-tuning

LoRA / QLoRA / DPO on your data. We start with the eval, never the loss curve.

03

Agents & tool use

Stateful, observable, recoverable. We ship agents that survive their second week.

04

Evals & LLMOps

Custom eval harnesses, cost dashboards, drift alerts. The boring part most teams skip.

05

AI-native product design

Surfaces, latency budgets, fallback UX. What does “loading” look like at 1.2 s of TTFT?

06

Strategy & advisory

Buy-vs-build, model selection, vendor lock-in. Hourly retainer for CTOs.

02 — Process

How an engagement actually runs.

  1. 01

    Week 0

    Discovery

    Two 90-minute calls. We map your data, your constraints, and your definition of success. Output: a one-page scope you can show your board.

  2. 02

    Week 1–3

    Prototype

    A working slice — not slides. End-to-end through your stack. We bring our own eval harness; you keep it.

  3. 03

    Week 4–8

    Productionise

    Hardening, observability, cost controls, on-call runbooks. By the end of week 8 your team is on duty and we’re paired with them.

  4. 04

    Week 9–12

    Handover

    Two weeks of pair-shipping. We document, we offboard, we leave the studio number with you for emergencies.

03 — Selected work

A field guide, not a deck.

Strato AI · Customer Operations · 2026

Cut first-response time from 6 hours to 47 seconds without firing a single agent.

A retrieval-grounded support agent trained on 12,840 resolved tickets. Hand-off to a human happens only on the 4% of conversations the model marks low-confidence — a number we instrument and tune weekly.

  • retrieval
  • brand-voice fine-tune
  • eval harness
  • human in the loop
Read the case study →

47s

p50 first response

↓ from 6h 12m

96%

CSAT after launch

↑ from 81%

$2.1M

Annualised savings

Year 1

4.1×

ROI on engagement

Validated Q1 2026

Northrun

Real-time pricing agent that recomputes 320k SKUs every 4 minutes.

Pricing · Retail

Cohostly

Multimodal listing QA — image + copy — caught 11% policy drift on day one.

Marketplace · Trust & Safety

Ink Foundry

Internal RAG over 4,200 PDFs of regulatory filings. Saved a $1.4M outside-counsel line item.

Legal · RAG

04 — Studio

A small team with deep scars.

LP

Lena Park

Founder · Applied research

ex-OpenAI alignment, ex-DeepMind

DR

Daniel Reyes

Principal · LLM engineering

ex-Anthropic, ex-Stripe ML

PS

Priya Shah

Principal · Product & design

ex-Figma, ex-Linear

MK

Marcus Kim

Studio · Operations

ex-Stripe COO staff

05 — Book

A 30-minute call,
no slides, no NDA.

You bring a problem you’re actually trying to solve. We bring a senior engineer and a clear yes-or-no by the end of the call.

Pick a slot → studio@synth.example

Discovery calls are free and confidential. We respond within one business day.