Production ML deployed since 2018

Models in production, not pilots in PowerPoint.

AI/ML development for mid-market companies that need real systems — not strategy decks. We build retrieval-augmented systems, fine-tuned models, and ML pipelines that ship to production and stay there. Postgres + Python + Ray. No "AI transformation" consulting.

Book a technical call → See top providers
47
production ML systems shipped since 2018
3mo
median time from kickoff to first production model
0%
junior data scientists on senior engagements
Shipped production ML for
TANBARK ROBOTICS Marlowe Research osprey.bio KIRBY FINANCIAL Quill Diagnostics caraway.energy
Why teams call us

Five patterns that send AI projects sideways.

AI/ML development engagements typically start with one of these. The discovery call is about which one is yours.

01

The data science team built a prototype that runs only on a laptop

It demos beautifully in Jupyter. Translating it into something that handles real traffic at 99.9% uptime turns into a six-month rebuild by a different team. Meanwhile leadership thought the model was "done" three months ago.

02

OpenAI/Anthropic spend doubled — and nobody knows which features drove it

There is no per-feature token tracking, no caching layer, no model router between cheap and expensive endpoints. The CFO wants a 40% cost reduction by quarter-end. The ML team needs three weeks just to instrument what is currently happening, before they can optimize anything.

03

RAG system hallucinates on questions about your own data

Vector search returns the wrong chunks half the time. Chunking strategy was decided in week one and never revisited. The retrieval evaluation harness exists in someone's notebook but has not been run since spring. Stakeholders are losing trust in the product because answers are sometimes confidently wrong.

04

Model drift killed accuracy and the monitoring fired late

Live performance degraded over six months but the dashboards only showed lagging metrics. Retraining is not automated, the feature store is half-built, and the data engineer who set up the original pipeline left in February. Each retrain takes two engineers two weeks.

05

EU AI Act audit found 30+ documentation gaps

Model cards exist for two of the eleven production models. Training data lineage is partial. Risk classification was never formalized. The compliance officer wants a remediation plan by next quarter and the technical answers do not exist yet — they have to be built before the documentation can be written.

What we build

Six AI/ML practices priced by deliverable.

AI/ML development priced by deliverable, not by hour. Senior ML engineers only — no entry-level data scientists on senior roles.

01

Production RAG systems

Retrieval-augmented generation deployed end-to-end — vector store selection, chunking strategy, retrieval evaluation harness, reranking, citation surfacing. Pinecone, Weaviate, pgvector, or Qdrant depending on scale. Evaluation suite ships with the system, not as an afterthought.
$95,000 – $320,000
02

Fine-tuning + model deployment

Open-weight model fine-tuning on your data (Llama, Mistral, Qwen), serving infrastructure on Modal/Replicate/in-house GPU, A/B routing between fine-tuned and prompted models. Includes evaluation framework against held-out data. We do NOT fine-tune when prompting works.
$120,000 – $450,000
03

LLM cost reduction

Per-feature token instrumentation, semantic caching, model routing (Haiku/GPT-4o-mini for cheap paths, frontier models for hard paths), prompt compression. Typical engagement reduces inference spend 40-70% within 8 weeks. We charge a flat fee; not a percentage of savings.
$45,000 – $140,000
04

ML platform + MLOps

Feature store, model registry, training/serving infrastructure on Ray or Kubeflow, monitoring (drift + accuracy + latency), automated retraining triggers. We build the platform once; your team operates it. Designed to be handed over, not rented from us forever.
$180,000 – $520,000
05

Multi-agent systems

Production agent systems with tool use, state management, observability, and bounded failure modes. LangGraph or custom orchestration depending on requirements. We focus on systems that handle real workloads — not autonomous-agent demos that fall over on edge cases.
$160,000 – $480,000
06

Senior ML engineer embedded

Staff or Principal-level ML engineer placed inside your team. Same standup, same Slack, same on-call rotation. Named individual on the contract. 6-12 month rolling engagements. We have done this with Series-B companies through Fortune 500 data teams.
$28,000 / month
How we run engagements

Four stages from data audit to handoff.

Same shape every time. Scope adjusts, the method does not. AI/ML development without method becomes science-fair work.

01.
Stage 01
Stage 01

Data audit

Two-week paid assessment. Inventory of your data, ML systems, and evaluation gaps. Output is a written report: what you have, what is missing, what is feasible at what cost. Flat $16,500 fee whether we proceed or not. Honest verdict on whether ML is even the right answer.

02.
Stage 02
Stage 02

Architecture

Written architecture decision records for every non-trivial choice — model selection, serving topology, monitoring strategy, retraining triggers. Includes evaluation methodology agreed before any model is trained. Compliance review with your team built in from day one.

03.
Stage 03
Stage 03

Engineering

Code in your repo from day one. Weekly demos to product and ops, not just data team. Evaluation suite ships with each iteration. Models go to production behind feature flags with rollback gates. PR review by your team is non-negotiable. Real workloads, not synthetic benchmarks.

04.
Stage 04
Stage 04

Handoff

Three-week parallel period per system. Your team operates the deployed model, we shadow. Runbooks current and tested. Retraining schedule documented. Drift monitoring alerting your on-call. 30-day and 90-day check-ins, then we are out and the system is yours.

Common questions

What we get asked before signing.

Questions from recent AI/ML development discovery calls. Honest answers, including disqualifying ones.

01 Why hire an AI/ML development firm instead of in-house data scientists?
For most companies, we recommend hiring in-house — eventually. We are a bridge during the period when you need senior ML capacity for a 6-18 month window but cannot recruit a Staff-level ML engineer in that timeframe (it currently takes 9-12 months). We typically build the first 2-3 production systems and design the platform, then your team takes over. We have helped clients hire 1-2 of our former engineers as full-time staff during handoff.
02 What if our data is not "AI-ready"?
It almost never is. The data audit phase explicitly checks this and a typical engagement allocates 30-40% of its budget to data engineering work that has to happen before any ML can ship. We do this work ourselves rather than handing it back to a "fix your data" recommendation. If your data is so far from usable that the cost-to-benefit math does not work, we will tell you on the discovery call.
03 Do you fine-tune models or just use OpenAI/Anthropic APIs?
Both, but mostly APIs. Fine-tuning makes sense when (a) inference cost is the bottleneck and you have over 100M tokens/month of traffic where a smaller fine-tuned model meets quality, or (b) your data has structure that no general model has seen and prompting consistently fails. For maybe 70% of engagements we use frontier models via API. We are happy to argue you out of fine-tuning if it does not fit your case.
04 How do you handle data privacy with LLM providers?
Zero data retention agreements with OpenAI, Anthropic, and Together via their enterprise plans. For regulated industries (healthcare, finance, EU) we deploy open-weight models on your VPC — Llama, Qwen, Mistral on Modal or your own GPU clusters. We have shipped HIPAA-aligned ML systems and EU AI Act compliant pipelines. Compliance reviews are baked into the architecture phase, not bolted on at the end.
05 What is your stance on multi-agent systems and "AI agents"?
Cautious. The agent demos that go viral are not the agent systems that survive in production. We build agents where the workflow has clear bounded states, evaluation is possible, and failures are recoverable. Open-ended "autonomous agent" systems that decide their own goals — we will tell you the engineering reality and usually steer toward simpler, evaluable architectures. Real production agents look more like state machines with LLM-powered transitions than the Twitter videos suggest.
04What gets shipped, what we charge

Five deliverables across a typical engagement.

Drawn from the last twelve completed AI/ML development engagements. Not the process — the actual artifacts you receive, with the band of cost industry observation suggests for each.

WRITTEN DATA AUDIT
A 14-page evaluation report
Inventory of data, ML systems, evaluation gaps, and feasibility scoring per use case. Includes the honest Go/No-Go recommendation on whether ML is even the right answer. Flat $16,500 fee whether the engagement proceeds.
ARCHITECTURE
Decision records + evaluation harness
Ratified architecture decision records — model selection, serving topology, monitoring strategy. Held-out evaluation suite shipped before any model is trained, so success criteria are agreed in writing first. Typically $32,000–$58,000 of the engagement.
FIRST PRODUCTION MODEL
Working system behind a feature flag
A primary model serving real production traffic, behind a flag with rollback gates. Drift monitoring wired into your on-call. Initial workload is typically 5-15% of traffic during the first month. Engineering cost band: $80,000–$220,000 depending on model complexity.
COST + DRIFT REPORT
Per-feature token attribution + drift dashboards
Per-endpoint, per-feature token tracking with cost attribution. Automated retraining triggers configured. Median reduction in monthly inference spend after this work: 40-65%. Cost band: $14,000–$28,000.
HANDOFF PACKAGE
Runbooks, retraining schedule, 90-day support
Operator runbooks tested under load. Documented retraining schedule. 30-day and 90-day check-ins after we are out. Final invoice tied to handoff completion — not calendar. Typically the last 8-12% of engagement spend.
05/What clients say

What our clients say.

"We had two failed prior engagements. The difference here was that they walked away from parts of the scope they could not own."

— Mei Silva, Head of Architecture leading 30 engineers at a B2B SaaS

"Mid-engagement scope change — they re-cut the backlog instead of running a change order. That kept the budget."

— Sarah Kim, Staff Engineer running an e-commerce platform

"Handoff included runbooks and a shadow period. Six months later we have not had to bring them back."

— Lucia Becker, Chief of Staff leading platform engineering

Tell us about your AI/ML situation.

Send the rough outline — current systems, what is breaking, what compliance is coming up, what your CTO is asking for. A senior ML engineer responds within one business day with questions or a direct next step.

Direct reply from a senior engineer
NDA before any technical or data specifics
If we are not the right fit, we say so on the call