Ai-Assisted Porting Throughput And Cost-To-Prototype

Issue 86 Edition 2026-03-27 5 min read

General

Sources: 1 • Confidence: Medium • Updated: 2026-03-27 10:08

Key takeaways

A first working Go version was built in about 7 hours with approximately $400 in LLM token spend.
The team used a one-week shadow deployment running old and new implementations in parallel to confirm behavior matched.
The case study frames the effort as having been completed in a day and as saving approximately $500K per year.
The author states that the "$500K/year saved" framing is somewhat hyperbolic.
A case study describes producing a custom Go implementation of the JSONata JSON expression language using an AI-assisted "vibe-porting" approach.

A first working Go version was built in about 7 hours with approximately $400 in LLM token spend.
A case study describes producing a custom Go implementation of the JSONata JSON expression language using an AI-assisted "vibe-porting" approach.

The team used a one-week shadow deployment running old and new implementations in parallel to confirm behavior matched.
The author attributes the speed of the AI-assisted port primarily to the existence of JSONata's test suite.

The case study frames the effort as having been completed in a day and as saving approximately $500K per year.
The author states that the "$500K/year saved" framing is somewhat hyperbolic.

What is the underlying cost model behind the claimed ~$500K/year savings (what costs were removed or reduced, and relative to what baseline)?
What was the total end-to-end engineering effort after the first working version (hardening, bug fixes, performance work, documentation, operationalization)?
How comprehensive was the JSONata test suite (coverage, edge cases, conformance targets), and did it require augmentation during the port?
What mismatches (if any) were found during the one-week shadow deployment, and how were they characterized (correctness vs. undefined behavior vs. performance differences)?
What were the production performance characteristics of the Go implementation relative to the prior implementation under real workloads?

AI assisted rewrites may reduce time and cost to reach a working prototype for language ports when a solid test suite exists, potentially shifting budgets from greenfield build to validation and hardening.
Shadow deployments running legacy and rewritten systems in parallel may become a standard de risking pattern, increasing adoption of rewrites in production environments.
ROI narratives around AI assisted engineering can be overstated without cost breakdowns, suggesting investors should discount headline savings claims until unit economics are demonstrated.

Multiple independent case studies show similar prototype speed and low token spend, with transparent reporting of total end to end effort including hardening and operations.
Documented shadow deployment results show minimal behavior mismatches and clear resolution paths, plus stable production performance relative to the prior implementation.
Savings claims are backed by a clear baseline and cost model, showing which expenses were reduced and how those savings persist over time.

Follow up disclosures show most effort occurred after the first working version, erasing the apparent speed and cost advantage when including testing, fixes, and operationalization.
Shadow deployments reveal frequent correctness gaps, undefined behavior differences, or performance regressions that require extensive manual debugging.
Cost savings claims fail to map to measurable reduced headcount or infrastructure spend, or are described as marketing framing rather than realized economics.