· verification
Verify every claim before you contact me.
An evidence dossier, not a portfolio. Every number below links to where you can check it, and each item says what kind of evidence it is. Where something cannot be checked publicly, it says so and why.
SCORECARD · every number, one place
run
three accuracy numbers, one run, measured three ways
- 94%
- retrieval hit-rate across all six suites - how often the right source document is among what the model reads, on the 253 cases that have one.
- 74%
- end-to-end pass rate across the same six suites - the whole answer graded, not just retrieval. A stricter bar, and the one that shows where the system still falls short.
- 96%
- FinanceBench retrieval hit-rate - the same retrieval test on the one external financial benchmark alone, where the end-to-end pass rate is 63%.
They measure different things, so they differ. All three - plus the 0/297 invented source citations - come from one benchmark run (2026-04-27), with methodology and per-case results published for audit.
· things I don't control
Third-party verification
Evidence produced or held by neutral parties - public benchmark datasets, funding databases, knowledge graphs, and credential issuers. Where I authored the record myself, the item says so.
Public benchmarks
public corpora · results auditable · runner closed
- claim
- Accuracy is graded mostly on public corpora I did not create, with the two internal suites named as internal.
- what this proves
- The per-suite scorecard and per-case results are published for audit. What is not public is stated: the runner that executes them is closed, so the runs can be audited but not re-executed.
- evidence
- 221/297 (74%) end-to-end pass rate and 239/253 (94%) retrieval hit-rate across six suites, run 2026-04-27 (Cohere
rerank-3.5) - 0/297 answers citing a document that was not retrieved
- 96% (144/150) retrieval hit-rate on PatronusAI FinanceBench (Islam et al., 2023, arXiv:2311.11944), where the end-to-end pass rate is 63% - the gap is over-abstention, explained in the case study
- Public corpora: CUAD (legal, NeurIPS 2021), FinanceBench, openFDA drug labels and SEC EDGAR 10-K filings. Internal: real-estate and consulting suites
- 221/297 (74%) end-to-end pass rate and 239/253 (94%) retrieval hit-rate across six suites, run 2026-04-27 (Cohere
- verify
- BENCHMARKS.mdraw runsFinanceBenchfull breakdown
Funded venture
third-party record
- claim
- Co-founded SculptAI (a 4-agent game-dev pipeline, 2024-2025) which raised $350K in seed funding.
- what this proves
- A third party committed capital. The seed round is on record in public funding databases; the amount is not publicly displayed there and is confirmed on request.
- verify
- Crunchbase: Definity LegendSculptAI case study
Knowledge-graph identity
public record · self-registered
- claim
- The "Nic Chin" entity is registered on Wikidata - the same identity layer LLMs use to ground entity claims.
- what this proves
- One disambiguated identity linking this site, GitHub and LinkedIn - not a name that could belong to anyone. The entry was created by me, so it proves who is who, not that anyone vouches for it.
- verify
- Wikidata Q138698158LinkedIn
Vendor credentials
issuer public · identifier on request
- claim
- Coursera course certificates issued by IBM (AI agents using RAG and LangChain), Microsoft (generative AI) and Google (prompting), plus a University of Northampton honours degree.
- what this proves
- The issuers are independent and named. These are course certificates, not professional certifications, and are described as such.
- verify
- identifiers on request
Independent client reviews
platform-verified · on request · NDA-restricted
- claim
- Upwork Top Rated Plus (in Upwork’s words, the top 3% of performers on the platform) with a 100% Job Success Score across multiple completed contracts.
- what this proves
- A neutral platform - not me - verified the review history and outcomes.
- verify
- shared in first call
· things I built
First-party evidence
Work I produced - but shipped in a form you can inspect yourself: live products you can use, open-source code you can read, a design you can audit, and the failures that produced the rules. That last one is first-party and not externally reproducible; it is here because omitting it would be the bigger tell.
Live products you can use now
first-party · publicly usable
- claim
- Two production SaaS products built and shipped as sole architect - publicly accessible, no demo videos in place of a live URL.
- what this proves
- The work runs in production, not just in a slide deck.
- evidence
- sureciteai.com - multi-tenant document intelligence RAG, free trial with no card
- systemaudit.dev - codebase intelligence reports, typically in under 3 minutes, free scan with no signup
- verify
- sureciteai.comsystemaudit.dev
Open-source code
first-party · openly inspectable · scope stated
- claim
- The SureCiteAI benchmark methodology and per-case run results are public. So is the SystemAudit repository scanner, under MIT - language detection, size and structure metrics, npm dependency extraction and health indicators.
- what this proves
- The published parts are inspectable - commit history, diffs, and raw artifacts included. What is not open is named rather than implied: the SureCiteAI eval runner, and SystemAudit’s security pattern rules and the enforcement pass that overrides the model, are closed.
- verify
- github.com/nicukBENCHMARKS.mdSystemAudit scanner (MIT)how the two layers split
Measured failures, and the rules they produced
first-party · measured · not externally reproducible
- claim
- Findings that were wrong, caught by measurement rather than by a user, with the standing rule each one produced.
- what this proves
- The system is measured rather than assumed. A project that cannot name what it got wrong has not been looked at closely enough to know.
- evidence
- SystemAudit pattern rules matched 13 candidates on a 41,000-line TypeScript codebase; 2 were real. Three guards followed - declaration-line, context-window, validator-presence. The number that mattered was the denominator, not the headline
- A clean scan implied full coverage it never had. Reports now state it: "Analysed 80 of 214 files" - an empty finding list means nothing found in what was read, never nothing wrong
- File-and-line evidence was produced by the deterministic layer and then stopped at the server. The architecture was right and the delivery was not - two separate defects, only one of them visible to users
- Exposed-credential locations were reachable by paying. They are now never gated, at any tier - a rule adopted because withholding them converts well, not because it does not
- verify
- the four changes in fullSystemAudit scanner (MIT)
A deterministic layer that overrides the model
first-party · design documented
- claim
- SystemAudit does not rely on LLM output alone - a deterministic pass re-checks every claim against measured facts and overrides the model where they disagree.
- what this proves
- A hallucinated finding cannot survive to the final report; correctness is governed, not trusted.
- verify
- the two-layer architecture in full
· the person
Engineering track record
The products are proven above. This is the record of the person behind them.
Engineering track record
first-party · externally corroborated
- claim
- Production AI systems architect and fractional AI CTO.
- what this proves
- The person behind the products has a public, checkable history.
- evidence
- 13 production AI systems designed and shipped
- Two live SaaS products as sole architect (SureCiteAI, SystemAudit)
- Published AI evaluation methodology with a public run history
- Fractional CTO engagements across funded startups
- Public technical writing and GitHub history
- verify
- About Nic ChinLinkedInGitHubPortfolio
· the limits
What is deliberately not here
- Named client testimonials with company logos. Engagements run under NDA.
- Screenshots of internal client systems, dashboards, or proprietary data.
- Full certificate URLs, which expose the holder's legal name. Issuer is shown; identifier is shared on request.
- The benchmark runner. The methodology and every per-case result are published, but the code that executes the runs lives in the private product repository, so the runs can be audited, not re-executed.
- A public link behind the AI citation count. Microsoft measures it and I cannot make their report public - it is private to the verified site owner. Unlike every benchmark above, you cannot audit that one yourself; I will screen-share it during procurement.
why this page exists
Most portfolios ask you to trust the author. This page is designed so you don't have to.
Wherever possible, claims are backed by public benchmarks, live software, open-source code, or third-party records anyone can inspect independently. Where something can't be verified publicly - client confidentiality, personal privacy - I say why, and provide verification during procurement.