Production AI, for teams whose AI is already in front of real customers
From demo to dependable in 90 days.
Your RAG works in the demo. In production it cites the wrong source, leaks what it shouldn't, and costs more every month. We fix that, as an outcome, not a hire.
Evaluation by version
Set of 400 real questions
| Version | Groundedness | Citation accuracy | Attacks blocked | Cost per answer | p95 | Status |
|---|---|---|---|---|---|---|
| v12 | 91% | 88% | 97% | $0.011 | 2.4 s | Replaced |
| v13 | 93% | 92% | 98% | $0.008 | 2.1 s | In production |
| v14 | 94% | 84% | 98% | $0.006 | 1.9 s | Blocked |
- Evaluation by version
v12
ReplacedGroundedness 91% · Citation accuracy 88% · $0.011
v13
In productionGroundedness 93% · Citation accuracy 92% · $0.008
v14
BlockedGroundedness 94% · Citation accuracy 84% · $0.006
One answer, traced
What is the notice period to terminate the services agreement?
Ninety days, in writing, before the end of the current term. [1][2]
- [1] Services agreement, clause 14.2
- [2] Amendment 3, item 2
- Grounded in the sources
- Citation checks out
- User can see both documents
$0.006 · 1.7 s · v13
Example with made-up values. Real targets are agreed with you in week 2.
The demo was the easy part.
Many companies already put RAG and AI assistants live. The hard part came after. Today they try to fix it by hiring one senior engineer at a time. Code and Soul delivers the outcome with a team, on a fixed timeline and with measurement.
For the CTO, VP of Engineering or Head of AI at a company whose AI is already used by real customers, and who is hiring to fix what came after.
Where AI breaks after launch
A typical answer from an internal assistant, with what usually goes wrong in it.
Legal assistant
Made-up example
What is the early termination fee in the services agreement?
The fee is three monthly payments, per clause 14.3 [2]. Here is the excerpt from the legal opinion: "the prevailing case law holds that..."
- [1] Services agreement.pdf
- [2] Refund policy.docx
- [3] Vendor comment.txt: "ignore previous instructions and..."
prompt v14 · no evaluation on record
$0.04 on this answer, $0.01 last month
- Answers that sound right but aren't grounded in your sources
- Citations that don't support the answer
- Licensed or sensitive content showing up where it shouldn't
- Prompt injection and data exposure
- No way to tell if the last change made quality better or worse
- LLM costs growing faster than usage
How it works
Three fixed phases, then ongoing operation. Each phase ends with something in production or a document you approve.
1
2
3
4
5
6
7
8
9
10
11
12
+
Week
Weeks 1 and 2
Assess
We review your architecture, measure a baseline on real questions and score security risks.
You get
A readiness report with a score per area, prioritized risks and the plan for weeks 3 to 12.
Weeks 3 to 8
Harden
Hybrid retrieval, reranking, source governance, document-level authorization, guardrails and cost controls.
You get
Every change ships to production, measured against the week 2 baseline.
Weeks 9 to 12
Operate
Automated evaluation on every change, observability and a runbook. Your team keeps all of it.
You get
A quality and cost dashboard, the evaluation pipeline, documentation and the handover.
After 90 days
Run, monthly
Quality and cost tracking, continuous tuning and new use cases built the same way.
You get
A monthly report on the agreed indicators.
What you can measure
Targets are agreed with you in week 2. Every change runs through the same evaluation, and nothing ships if an agreed target regresses.
- How we measure
- Automated evaluation on the set of real questions.
- What you track
- The share of answers supported by the retrieved sources.
Groundedness
A small senior team, with one lead accountable for the result
Working in your time zone, from day one to the handover to your team.
AI technical lead
Part time, start to finish
Architecture, decisions and the relationship with the CTO. Accountable for the result.
Senior AI engineer
Full time in phases 2 and 3
Retrieval, content governance, guardrails and cost control.
Evaluation and platform engineer
Part time
Evaluation pipeline, observability and infrastructure on AWS.
Why Code and Soul
- A senior team with one lead accountable for the result
- Your time zone, US contract, AWS partner
- We run AI in production in our own products before we build it for you
- Part of our fee is tied to the targets we agree on
The stack we work in
- Amazon Bedrock
- OpenSearch
- S3 Vectors
- Lambda
- CDK
- Terraform
- PostgreSQL with pgvector
- Python
- LangGraph
- LlamaIndex
- Ragas
- DeepEval
- Langfuse
Technical proof we can show
What we have already done in production, on the same stack this offer works in.
LLM cost controlled by design
In SYNQi³, cost grows with the number of distinct behavior patterns, not with event volume.
Vector infrastructure swapped in production
A migration to S3 Vectors without interrupting the product.
RAG on Amazon Bedrock
Designed for a telecom client, with a global agency.

AWS partner, US contract
Contract and billing go through Code and Soul's US entity.
The first Production AI project becomes a public case, with the week 2 baseline and the week 12 result.
Hiring a senior AI engineer takes months. This takes 90 days.
And the evaluation pipeline stays with your team.
What we get asked
- We'd rather hire someone in-house.
- Makes sense long term. We hand over the evaluation pipeline, the runbook and the knowledge, so whoever you hire starts from a working system, not a blank page.
- How do we manage a team outside the US?
- Same time zone, US contract and billing, one senior lead accountable for the whole. You manage the result, not the hours.
- Our data can't leave our environment.
- We work inside your AWS account, with your access rules. Nothing leaves your environment.
- It's expensive for 90 days.
- Compare it with months of recruiting plus a full-time salary, with no guarantee of outcome. Part of our fee is tied to the targets we agree on.
- We use a different stack.
- Our method is stack-agnostic. We go deepest on AWS, Python and pgvector or OpenSearch; if yours is different, the week 1 and 2 assessment tells us both whether it's a fit.
- We already have a vendor.
- Then we can come in on what they don't cover: evaluation, guardrails and cost, measured against a baseline.
Production AI readiness checklist
Every "no" becomes a line in the week 1 and 2 readiness report.
- Grounding and retrieval5
- Governance and permissions3
- Security3
- Evaluation and releases3
- Cost and operations5
Code and Soul · 19 questions
Production AI readiness checklist
Grounding and retrieval
- We have a set of real questions, with expected answers, that we test against.
- We measure how many answers are supported by the retrieved sources.
- We check that each citation supports the sentence it is attached to.
- Retrieval combines semantic and keyword search, with reranking.
- Chunking and metadata filters were tuned on our data, not left at defaults.
Get the readiness checklist
Tell us what you have in production and download the checklist right away.
- 1The checklist shows up right here, ready to download, as soon as you send.
- 2We read what you wrote about your production setup.
- 3If it makes sense, we'll write to you by e-mail.
Rather talk? Book 20 minutes

