Humaniti AI — Evidence-Grounded Classroom Practice Analysis
Privacy-first multimodal instructional assessment engine assigning deterministic pedagogical ratings grounded exclusively in verifiable, timestamped evidence
The Problem & Target Context
Automated educational assessment tools routinely hallucinate subjective judgments, penalize missing data as incompetence, and utilize biased demographic proxies such as vocal cadence, accent, or facial emotion.
Engineering Approach & Core Principles
Engineered an evidence-grounded assessment platform enforcing strict ethical invariants: every finding links directly to exact timestamp offsets, missing evidence is labeled not_assessable rather than zero, and grades are withheld if less than 70% of the rubric is observable.
What I Built & Individual Contributions
Architected the Next.js 16.4 App Router web client, PostgreSQL database schema with Prisma ORM, ephemeral media processing pipelines, deterministic 4-tier rubric scorer (Poor, Average, Good, Excellent), and anti-bias evaluation filters.
Architecture & Data Pipeline
Classroom Audio/Video → Ephemeral Media Stager (.private/) → Whisper Transcription & Audio Diagnostics → Evidence Grounding Engine → Deterministic 4-Tier Classification → Coaching Report
[ Classroom Media Recording ]
│ (Local Ephemeral Staging)
▼
┌────────────────────────────────────────────────────────┐
│ Humaniti Grounded AI Engine │
│ ├── Audio & Transcript Diagnostic Pipeline │
│ ├── Timestamped Evidence Item Locator (start/endSec) │
│ ├── Anti-Bias Filters (No Accent/Emotion Proxies) │
│ └── Fail-Safe Grade Withholder (<70% Rubric Weight) │
└───────────────────────────┬────────────────────────────┘
│
▼
┌────────────────────────────────────────────────────────┐
│ Next.js & Prisma Coaching Hub │
│ - Deterministic 4-Tier Rating (Poor -> Excellent) │
│ - Transparent Auditable Video Timeline & Quotes │
└────────────────────────────────────────────────────────┘Key Engineering Decisions & Trade-Offs
Enforced strict evidence-first invariants: every score requires verified startSec/endSec timestamp locators. Missing observations are marked not_assessable rather than zero. Automatically withholds pedagogical classifications ('insufficient_evidence') if observable rubric coverage falls below 70%.
Testing, Verification & Reproducibility
Full test suite with mock audio and transcript fixtures. Ephemeral processing scrubbers verified via automated cleanup tests. Synthetic zero-dependency demo mode accessible via '/demo' route.
Verified Outcomes & Metrics
Zero PII retention policy; 100% timestamped evidence attribution; automatic grade withholding under sparse observations; WCAG 2.1 AA accessible UI.
Verified Implementation
Inspected repository code enforces ephemeral scrubbing of media upon job completion, includes synthetic zero-dependency demo fixtures, and implements strict Zod schema validation across all assessment endpoints. Verification here represents code audit and local inspection, distinguishing functional code paths from external cloud provider accounts.
Boundaries & Known Limitations
Does not claim objective measurement of lifelong educator quality. Designed exclusively for structured coaching feedback on observable classroom practices; cloud LLM provider calls remain blocked by default.
Related Portfolio Projects
Explore companion systems in similar domains across AI engineering, quantitative finance, and developer infrastructure:
Collaborating on AI Engineering or Quantitative Systems?
I design and build end-to-end production software, quantitative models, and reliable AI architectures. Let’s discuss technical challenges or product collaboration.