Time to accepted change
Does faster generation survive review and rework?
Trace a sample from task to merge: cycle time, review wait, revision rounds, and developer effort. Compare similar work where the data allows.
AI ENGINEERING EFFECTIVENESSVELOCITY · QUALITY · COST
Make AI coding
work for your team.
Measure and improve AI-assisted development across delivery speed, code quality, and cost. Start with a focused engineering audit.
For engineering teams using Cursor, Claude Code, Codex, or Copilot.
Understand what changes after the code is generated.
✳ WHAT THIS ISMore code.
More code.
But better delivery?
AI adoption is only part of the picture. We examine the work around it: review, rework, architecture, and the cost of getting a change accepted. Then we help improve the workflow.
01 / HOW YOU STARTOne team. One workflow.
One team. One workflow.
A clear first assessment.
For established software teams already using AI coding tools. A fixed scope and fee, agreed before we start.
AI ENGINEERING EFFICIENCY AUDIT
Find the friction.
Improve the workflow.
Start with one team, one repository, and a representative sample of recent changes. Review tooling spend, agent setup, and the path from task to accepted code.
Baseline, prioritized findings, and one agreed workflow improvement. Larger implementation is scoped separately.01
Establish the baseline.
Review available usage exports, recent PRs, CI history, and developer interviews. Identify what can be measured and where the evidence is incomplete.
BASELINE & EVIDENCE GAPS02
Diagnose the tradeoffs.
Trace review burden, retries, context use, and model choices. Rank improvements by likely value, effort, and confidence.
PRIORITIZED FINDINGS03
Make a bounded change.
Implement one agreed change to agent instructions, context, model selection, or review checks. Leave a repeatable evaluation and a follow-up measurement plan.
CONFIGURATION, RUNBOOK & NEXT STEPS02 / AUDIT SCOPELook beyond
Look beyond
the completion count.
Four connected areas. We investigate the parts that matter to your team, using the evidence you actually have.
One engineering workflow.
Four connected perspectives.
VELOCITY WITH REVIEW INCLUDED
01Task
02Generated change
03Review & rework
↗Accepted work
Examine where AI shortens a task and where it shifts effort onto reviewers. Look at task boundaries, PR size, revision loops, and the checks needed before a change is accepted.
- +Task-to-merge workflow baseline
- +Review effort and retry analysis
- +PR scope and handoff recommendations
- +A plan to measure comparable work
03 / WHAT WE MEASUREEvidence you can
Evidence you can
make decisions with.
These are assessment questions, not claimed client results. We establish a baseline before recommending a change.
Code your team can maintain
What does the new code leave behind?
Inspect PR size, duplication, abstraction reuse, churn, and CI failures. Review the code and its context rather than treating line counts as a quality score.
The cost of accepted work
What are you paying for a usable result?
Combine available tool and model spend with retries and estimated review effort. Separate measured costs from estimates and account for missing telemetry.
PR history alone cannot reliably identify AI-written code or prove that AI caused a change in performance. We combine available tool data, code review, and team interviews, and make uncertainty explicit.
04 / THE COMPANYClose to the code.
Who we are
Close to the code.
Responsible for
the outcome.
You work directly with Raffay Sajjad: the engineer who assesses your workflow and implements the agreed changes.
Production engineering, agentic development, and model economics in the same conversation.
One engineer, from assessment to implementation.From the first question through handover.
BACKGROUND
01 / LEADERSHIP
Built and led engineering teams.
Technical leadership at Trafilea across architecture, delivery, experiments, and production systems.
02 / PRODUCT
Shipped a product from scratch.
Finly AI: mobile, backend, infrastructure, and analytics. Idea through a working product.
Product engineering03 / AI PRACTICE
Inside the AI development workflow.
Hands-on practice with coding agents, context management, model selection, and local models. The focus is accepted, maintainable work.
Founder experience behind Tiercel Labs. Not client logos.
WORKFLOWS WE ASSESSCursorClaude CodeCodexGitHub CopilotCustom agents & local models
05 / WRITINGEngineering practice.
Engineering practice.
Under examination.
Selected writing on engineering productivity, review burden, and the economics of AI-assisted work.
All 28 articles
Lab speed is not field pull requests.
A controlled Copilot task was 55.8% faster. Field experiments at Microsoft and Accenture show smaller, noisier PR gains.
ReadThe cost of retries and review.
Total cost is inference plus retries plus human review plus delay. A cheap model can still be expensive.
ReadMeasure the work. Not the tokens.
Token price is easy to compare. Cost per accepted task is what decides whether a model choice is cheap or expensive.
ReadQUESTIONS
Straight answers.
01Who is this for?
CTOs, VPs of Engineering, engineering directors, and platform or developer-experience leads at established software teams already using AI coding tools. You have a real workflow to examine and an owner who can help implement changes.
02What does the first audit cover?
One team, one repository, and a representative sample of recent work. We agree the questions, available evidence, deliverables, timeline, and fixed fee before starting. A focused assessment typically takes one to two weeks after access is ready.
03What do we receive?
An evidence-backed baseline with its limitations, prioritized findings, one agreed workflow or configuration improvement, and a runbook for measuring what happens next. Larger implementation work is a separate decision.
04What access do you need?
Usually usage or billing exports, selected PRs and CI history, repository instructions, and conversations with a few developers and reviewers. We agree the minimum access needed. Redacted exports or a guided walkthrough can be used when direct repository access is unsuitable.
05Can you guarantee savings or prove AI wrote a change?
No fixed saving is promised before measurement. Repository history alone does not reliably establish AI authorship or causality. We distinguish observed data from estimates and use comparable tasks where possible. Faster generation is not treated as proof of faster delivery.
06Do you implement the recommendations?
Yes. The first audit includes one bounded improvement agreed in the scope, such as repository instructions, context configuration, or a review check. Broader tooling and workflow changes are scoped separately after you review the findings.
07How much does it cost?
The first engagement has a fixed fee based on the repository, questions, and access available. We send a written scope and quote before you commit. There is no ongoing retainer required.
08Do we need a new tool or a particular vendor?
No. We start with your existing tools and available data. Cursor, Claude Code, Codex, GitHub Copilot, custom agents, and local models can all be part of the assessment. Tool names describe the workflows we assess, not vendor partnerships.
NEXT STEPSTART WITH YOUR ENGINEERING WORKFLOW.
More AI-generated code.
Better engineering?
Tell us which coding tools your team uses and where cost, review, or delivery is becoming difficult to understand.