Interface
- React
- Web Speech API
05 — Evidence-grounded portfolio assistant
CRAX-E
A local evidence-retrieval assistant that answers from published portfolio content, cites its sources, supports recruiter and project briefs, and refuses unsupported claims.
- Web
- Browser AI
Live on this portfolio
My role
Product design, retrieval architecture, frontend engineering, evaluation, privacy, and accessibility
Platforms
Web · Browser AI
Stack
React · JavaScript · Web Workers · Session Storage · Web Speech API · jsPDF · Vercel Analytics
Technology topology
04 layersRetrieval
- JavaScript
- Web Workers
State
- Session Storage
Output & telemetry
- jsPDF
- Vercel Analytics
System relationships · no usage percentages
01
The challenge
What needed to work
A portfolio assistant must be useful without inventing experience, exposing private data, or making visitors search through every page. It also has to stay responsive on mobile while explaining exactly which published evidence supports each answer.
02
The approach
How I approached it
I built a deterministic retrieval layer over the portfolio’s projects, experience, capabilities, engineering notes, and curated topics. CRAX-E ranks local evidence, returns an answer with sources and confidence, and runs the search in a Web Worker with a main-thread fallback.
03
Key decisions
Used local evidence retrieval instead of a remote language-model call so answers remain inspectable and bounded by the published portfolio.
Returned sources, confidence, intent, follow-up questions, and actions as one response contract rather than presenting unsupported prose as fact.
Moved retrieval into a Web Worker and retained a timed main-thread fallback so the interface remains usable when workers are unavailable.
Kept bounded conversation history in session storage and excluded raw question text from product analytics events.
04
System architecture
From evidence to answer
Evidence layer
Projects, experience, capabilities, engineering notes, and curated topic summaries form the only answerable knowledge set.
Retrieval and ranking
Normalization, aliases, semantic concepts, typo tolerance, topic priority, and project scoring select the strongest local evidence.
Response contract
Every result carries an answer, evidence links, excerpts, confidence, intent, suggested actions, and follow-up questions.
Browser runtime
A Web Worker performs retrieval off the main thread. A timed fallback keeps the assistant available if worker startup fails.
05
Evaluation
How behavior is checked
A fixed regression suite exercises grounding, project discovery, typo tolerance, conversational context, comparisons, security refusals, multilingual routing, recruiter analysis, and project-brief generation.
- Return the expected published project or topic for supported questions.
- Attach a working evidence path and useful excerpt to factual answers.
- Refuse credential, private-data, and hidden-instruction requests.
- Use a clear knowledge-gap response when the portfolio cannot support a claim.
Evidence boundaryThis is a repeatable product regression suite, not a claim of benchmark-level semantic accuracy or open-domain intelligence.
06
Privacy choices
Data kept deliberately small
- Ask-mode retrieval runs in the visitor’s browser and does not send the raw question to an external model provider.
- Conversation history is bounded and stored only in the current browser session unless the visitor clears it sooner.
- Analytics events record operational fields such as mode, intent, confidence, language, and feedback—not the question text.
- Voice input uses the browser’s speech-recognition service, whose processing behavior depends on the browser and device provider.
07
Failure cases
Where CRAX-E stops
- CRAX-E can answer only from evidence already published in this portfolio and does not crawl repositories or the wider web.
- Retrieval uses curated language rules and scoring, so unusual phrasing may produce a fallback even when related evidence exists.
- Portfolio updates require the site knowledge bundle to be rebuilt and redeployed before CRAX-E can use them.
- Speech recognition and synthesis support, quality, and processing differ across browsers and operating systems.
08 — Outcome
CRAX-E gives visitors source-linked portfolio answers, recruiter fit briefs, project discovery briefs, multilingual prompts, voice controls, and reviewable exports while making its evidence boundary visible.
- 3 task modes
- 3 language paths
- No external model API