Talent Evidence OS

Inspiration

Talent decisions often happen with messy evidence: scattered project updates, chat excerpts, manager impressions, delivery records, and incomplete feedback. The problem is not that teams lack information. The problem is that they often mix strong evidence, weak evidence, and personal impressions without knowing which is which.

Talent Evidence OS was inspired by a simple question: what if AI could help teams reason about people more responsibly, by making the evidence boundary explicit instead of pretending to be an all-knowing judge?

What it does

Talent Evidence OS turns work-related evidence into a structured, auditable talent assessment.

Users provide a business goal, target role, evaluation period, decision purpose, and evidence list. The system classifies evidence quality, separates performance from potential, identifies missing evidence, flags unsafe or unauthorized data, and produces a human-review-ready report.

It does not make final hiring, firing, promotion, compensation, or layoff decisions. Instead, it helps teams understand what can be concluded, what cannot be concluded, and what should be validated over the next 60-90 days.

How we built it

We built Talent Evidence OS with Codex and GPT-5.6 as an agentic workflow.

Codex helped turn the talent-evaluation framework into a reproducible project structure, sample data, prompts, report templates, and local validation flow. GPT-5.6 powers the reasoning layer: evidence classification, role-fit analysis, confidence calibration, bias checks, missing-evidence detection, and action-plan generation.

The project uses synthetic or anonymized sample data so the full workflow can be tested without exposing private employee information.

Challenges we ran into

The biggest challenge was keeping the system useful without making it unsafe.

Talent evaluation is high-impact. A naive AI tool could easily over-score people, rank employees from weak evidence, or turn private chat behavior into permanent labels. We had to design strict boundaries: reject unauthorized data, avoid final employment decisions, require evidence quality checks, and make human review part of the output.

Another challenge was making the output practical. Managers do not need a vague essay. They need a clear view of evidence strength, risk, missing proof, and what to do next.

Accomplishments that we're proud of

We are proud that Talent Evidence OS treats uncertainty as a first-class feature.

Instead of forcing a false answer, it can say: evidence insufficient, confidence low, collect these data points, run this 60-90 day validation plan, and escalate this decision to HR or leadership review.

We are also proud of the evidence model. It separates verifiable work results, structured feedback, concrete incidents, and vague impressions, so the user can see why a conclusion is strong, weak, or not supported at all.

What we learned

We learned that responsible AI for talent work is less about replacing judgment and more about improving the quality of judgment.

The most valuable part of the system is not a score. It is the discipline of asking: What role are we evaluating against? What evidence is actually verifiable? What data should not be used? What decision is too high-impact for automation? What would change our mind in the next 90 days?

What's next for Talent Evidence OS

Next, we want to add a cleaner web interface, richer sample datasets, exportable PDF/HTML reports, and team-level organization diagnosis.

We also want to add integrations for structured work evidence such as project trackers, review docs, and calibrated interview notes, while keeping privacy, consent, and human review at the center of the product.

Built With

• cloudflare-pages
• codex
• evidence-calibration
• gpt-5.6
• hr-tech
• next.js
• typescript