Production system • Platform case study

Evidence-First Engineering Knowledge System

A static engineering publication paired with a grounded assistant that answers only from reviewed public evidence.

Try Ask Mantosh Inspect source

This page answers:How do you make an AI-assisted engineering portfolio useful without letting its claims outrun public evidence?

Problem

A portfolio can make claims without making them easy to verify. A generic chatbot creates the opposite risk: confident answers that outrun the evidence. This system was designed to make engineering work fast to inspect while keeping every AI answer inside a public, reviewable knowledge boundary.

Challenge

The system had to connect static publishing, knowledge synchronization, hybrid retrieval, generated responses, citations, quotas, and release controls without placing credentials in the browser. It also had to remain conversational while failing safely whenever evidence was missing.

My Contribution

I designed, implemented, tested, documented, and operate the complete public system: the website, knowledge model, indexing path, Worker, retrieval policy, Ask Mantosh experience, evaluation suite, and release workflow.

Architecture

Two controlled paths, one public evidence boundaryReviewed knowledge is indexed separately from visitor traffic; answers remain grounded in the same public source.
  • Public browserContent and questions; no credentials
  • GitHub control planeReviewed source, release gates, and OIDC identity
  • Cloudflare runtimeQuotas, retrieval, indexing, and model access

Publish and index

Only reviewed public knowledge enters retrieval.

  1. GitHubReviewed MarkdownPublic source of truth
  2. GitHubRelease gatesContent and safety checks
  3. BoundaryOIDC syncShort-lived identity
  4. RuntimeIndex WorkerApproved content only
  5. RuntimeD1 + VectorizeExact and semantic index

Ask and answer

Every generated response must trace back to retrieved evidence.

  1. BrowserVisitor questionBounded public input
  2. RuntimeWorker gateValidate, limit, or route
  3. RuntimeHybrid retrievalExact + semantic results
  4. RuntimeWorkers AISelected context only
  5. BrowserCited responseApproved source links

Evidence boundaryNo browser secret. No answer from model memory when public retrieval is insufficient.

Components and controls

  • GitHub Pages serves the static portfolio, project case studies, Insights, résumé, newsletter, and accessible interaction layer.
  • Markdown in Git is the source of truth for Ask Mantosh. GitHub Actions authenticates indexing with short-lived OIDC credentials.
  • Cloudflare Worker validates requests, enforces quotas, handles deterministic greetings and navigation, and coordinates retrieval.
  • D1 FTS5 and Vectorize provide lexical and semantic retrieval; reciprocal-rank fusion combines the results.
  • Workers AI receives only selected public evidence and returns structured answers with approved citations.

Engineering Decisions

Require evidence before generation

I chose retrieval confidence as the gate for generated answers because a polished unsupported response would weaken the entire portfolio. When public evidence is insufficient, the assistant declines instead of falling back to model memory. This reduces answer coverage, but makes every supported response inspectable.

Route deterministic work before inference

Greetings, navigation, public profile facts, and clear scope boundaries do not need semantic retrieval or generation. Handling them deterministically improves speed, preserves the AI allowance, and removes avoidable model variation. The trade-off is maintaining a narrow routing layer without letting it become a brittle keyword chatbot.

Separate publication and runtime trust boundaries

I kept reviewed knowledge in Git and used short-lived deployment identity for synchronization rather than exposing an indexing credential to the browser. The public client can ask questions but cannot modify knowledge or access the model directly. This adds deployment machinery, but keeps authority aligned with the risk of each operation.

Make release confidence executable

I treated content quality, accessibility, assistant behavior, performance budgets, and production health as one release problem. Automated gates and revision-aware smoke checks catch regressions before or immediately after publication. The suite costs maintenance time, but replaces informal confidence with repeatable evidence.

Outcome

Before

Portfolio claims, technical content, navigation, and AI answers could be evaluated separately.

After

One reviewed knowledge system connects published evidence, grounded answers, discoverability, and automated release confidence.

Evidence

A versioned offline suite exercises the production Worker entry point with controlled D1, Vectorize, Workers AI, and rate-limiter bindings.

Public implementationSource, configuration, knowledge documents, and workflows are inspectable in GitHub.
Evaluation record554 labelled cases and 9,734 objective assertions currently pass.
Release gatesStatic audits, Worker tests, browser journeys, accessibility scans, budgets, and smoke checks run on publication.
Architecture diagramThe public diagram above exposes data paths and trust boundaries without hiding the system behind prose.

Boundary: controlled fixtures do not measure live Vectorize recall, human preference, user satisfaction, or organizational impact.

Inspect the evaluation method and results

Limits

The implementation, public knowledge corpus, automated tests, release gates, and production-state documentation are available in the repository for inspection. The suite runs 22 browser checks across 11 visitor journeys in desktop and mobile Chromium; failure artifacts support review without pretending screenshots are pixel-perfect approvals.

Current limits: GitHub Pages rather than a custom domain, Cloudflare free-tier quotas, no declared staging environment, formal accessibility conformance audit, human-rated or live-production retrieval benchmark, cross-browser automation, or production Core Web Vitals trace.

Inspect the release-confidence method

What This Project Demonstrates

Evidence-Bound AI Engineering

Designed generation around reviewed sources, explicit confidence, approved citations, and safe refusal.

Platform Architecture

Connected publishing, indexing, retrieval, conversation, and navigation as one maintainable system.

Trust-Boundary Design

Separated public clients, deployment identity, stored knowledge, and model access with fail-closed controls.

Release Engineering

Turned content, accessibility, performance, assistant behavior, and production health into automated release gates.

User-Centered Systems Thinking

Adapted one evidence base for different audiences while keeping discovery fast and claims consistent.

This project demonstrates my ability to turn trustworthy engineering knowledge into a tested, secure, and usable platform capability.

Reflection

I would keep the evidence-first boundary and deterministic routing because they make the system useful without turning the portfolio into an ungrounded chatbot. With today’s knowledge, I would introduce a staging environment and human-rated, live-retrieval evaluation earlier; the current controlled suite is strong at regression detection but does not measure real visitor preference or production recall.

Need a platform that turns scattered knowledge into a dependable workflow?

View experience Discuss the constraint