Evidence saved
Recruiting email, resume versions, corrections, and profile facts enter as typed source records.
Loading
Preparing your workspace.
Loading
Preparing your workspace.
Product architecture
Technical architecture
Build is the technical landing page: source evidence comes first, bounded model reasoning handles semantic judgment, and risky output stops before it can write application or reminder state.
The hosted product uses the production stack below. The public CareerOS repo is a limited judge-safe Next.js demo/source package for the same evidence -> extraction -> review gate -> state loop.
Agent trace
Input
owner, timestamp, source id, redacted evidence
Agent
bounded prompt, structured output, confidence
Gate
review reason, fallback path, blocked mutation
Write
idempotent update, audit row, visible trace
Recruiting email, resume versions, corrections, and profile facts enter as typed source records.
Small agents triage, extract workflow facts, identify contacts, check risk, and plan reminders.
Low-confidence or risky changes pause with source evidence before application state mutates.
Provider, model path, purpose, confidence, evidence source, fallback, and review outcome stay visible.
Pipeline eval proof
This is not a broad ML benchmark. It proves the submission loop: bounded evidence enters the pipeline, extraction proposes application facts, risky output hits a review gate, and approved results become application or reminder state. The current Kaggle/Gemma proof path uses `gemma4:31b` via Ollama Cloud; the production product is intentionally task-routed and provider-configurable so fast triage, structured extraction, review evidence, resume reasoning, and artifact analysis can use different models behind the same review contract.
Eval result
15/15All fixtures passed action, stage, gate, and safety checks.
Apply, review, and ignore decisions matched expected outcomes.
Recruiting cases mapped to the expected application stage.
Risky or ambiguous cases paused before trusted writes.
High-stakes cases did not silently mutate state.
Task-routed models
The product contract is not tied to one model tag. Each task keeps its own purpose, structured output, confidence threshold, fallback path, and review gate, so better or cheaper models can be swapped per job without changing the user-facing evidence trail.
Fast relevance route
Cheap filtering decides whether a message belongs in the recruiting pipeline before heavier extraction runs.
Structured-output route
Company, role, stage, deadline, contact, and next action are parsed into bounded DTOs.
Higher-reasoning route
Risky or ambiguous updates are checked against source evidence before the write gate opens.
Long-context route
Resume feedback can use stronger reasoning over candidate history without changing inbox sync behavior.
Document or multimodal-capable route
Offer PDFs, interview screenshots, and uploaded evidence stay source-linked and reviewable.
Seven-stage pipeline
The product is intentionally split into narrow stages so every output has a source, a confidence level, and a review fallback.
Message ids, thread links, source snippets
Recruiting mail versus inbox noise
Company, role, deadline, status, contact
Scam, weak match, low confidence
Approve, correct, or dismiss
Application event plus reminder
Confirmed corrections become reusable context
Architecture overview
Public site, authenticated workspace, signed proxy
Application services, review gates, audit contracts
Applications, inbox evidence, reminders, memory
Fast triage, structured extraction, review evidence, resume reasoning, and artifact analysis can use different configured models
Runtime signals