I build agent systems that check their own work, and I ship the evidence with them.
Most AI demos prove a system succeeds on a case it was built for. That is the easy half. The half that decides whether you can leave something running is what happens when it is wrong.
Evaluation harnesses, not demos. Every system I deliver ships with a way to measure whether it actually works, on a held-out set, with the number visible. When someone later changes a chunk size or swaps a model, you find out whether it got better instead of guessing.
Validation loops that bounce. Agents propose, a separate checker runs the code rather than reading the claim about it, and work goes back until it converges. A loop that never bounces has not been shown to work. It has only been shown to run.
Production data integration. Eight years of it. Connectors for SAP ODP, Salesforce, Salesforce Marketing Cloud, ServiceNow, Workday, NetSuite, OData and the Google stack, landing into Snowflake, Databricks, Redshift, BigQuery and Synapse. If your data lives somewhere awkward, I have probably already moved it.
| Pull requests merged | 329 |
| Pull requests reviewed | 439, for ~44 engineers across 19 repositories |
| Tickets delivered | 569, ~97% closed |
| Production incidents resolved | 107, none left open, 34 at critical severity |
| Contributions in the last year | 1,072 |
agent-eval-demos — three small demos on verifiable evaluation. Each runs from a clean checkout with no API key, no GPU and no network, and each ships the number it claims alongside the command that produces it.
rag-evaluation-demo— retrieval with the harness attached, not bolted on.6/6held-out, 7 tests.validation-loop-demo— rejects a plausible-but-wrong first attempt, accepts the corrected second.brief-to-report-demo— structured input to a reviewable artefact, end to end.
Measurement work in the open: I read
PrimeIntellect-ai/community-environments
end to end and found that CHANGED_ENVS=none collapses their pytest parametrisation to an empty
set, so a run reports 6 skipped, exits 0, and the badge goes green. Their pytest major-version
upgrade was validated by a run that executed none of their tests. Every claim in that issue ships
with the command that produced it, so any of it can be re-derived or thrown out.
I state the number and its caveat. A 6/6 on a six-item fixture is a demo, not a benchmark, and
I say so in the README rather than letting it be quoted as one.
I answer no where the answer is no. Unverified is a professional answer; a confident guess is not.
I would rather tell you the build is the wrong build before taking money for it.
Java · Python · TypeScript · React Manchester, UK