Kaggle Gemma 4 Good judge demo

CareerOS public demo: recruiting mail becomes a reviewed job pipeline

This route runs a sanitized fixture through evidence, extraction, and review boundaries without Gmail or a model key. Optional Gemma readiness is reported separately from the deterministic fixture path.

It is not the full hosted Other Candidate source; it is the public demo/source repo for the agentic recruiting pipeline.

random recruiter email with missing contextspreadsheet drift after 50-100 applicationsstale follow-up reminders when the process movesJD, resume version, source, contact, and notes tracked together

Open review gate Connect Gmail sync Set up Gemma Inspect pipeline JSON

Agent handoff

Mailbox to pipeline state

Mailbox triage agent

Classifies recruiter mail, OA invites, follow-ups, rejections, offers, and vague emails with missing context.

intent: recruiter_oa_follow_up

Workflow extraction agent

Maps the thread to company, role, source, JD link, resume version, recruiter contact, deadline, and action.

proposal: update_application

Evidence/review agent

Attaches bounded evidence snippets and blocks low-confidence updates behind explicit review.

review_gate: required

Resume/context agent

Adds candidate context so the system can prioritize replies, fit, and preparation work.

context: candidate_ready

Reminder/notification agent

Creates deadline and follow-up notifications after review, then suppresses stale reminders once the thread moves.

notify: oa_due_no_stale_followup

Model router/provider layer

Prefers Gemma via Ollama Cloud when the API key is ready, falls back to deterministic parsing when unavailable.

provider: gemma_or_rules

Sanitized fixture

Recruiter evidence becomes a reviewed tracker proposal

real paincandidate note

What role is this recruiter email for?

I have 83 applications in a spreadsheet and cannot tell which role this OA belongs to.

recruiter emailmira.chen@heliosdata.example

Helios Data online assessment reminder

Your OA for the Machine Learning Platform Intern role is due Friday at 5:00 PM PT. Use the JD link from Handshake if you need the role details.

evidencemira.chen@heliosdata.example

Re: Helios Data online assessment reminder

You applied with resume-ml-v4.pdf. Reply here when the OA is complete and I can confirm the next interview window.

01Email evidencerandom recruiter thread

02Structured proposalfields + confidence

03Review gateaccept, correct, or dismiss

04Tracker statedeadline + notification

Extracted application proposal

Review required

CompanyHelios Data

RoleMachine Learning Platform Intern

SourceHandshake

JD linkhelios.example/jobs/ml-platform-intern

Resume versionresume-ml-v4.pdf

Recruiter contactMira Chen

DeadlineMay 15, 5:00 PM PT

Next actionComplete OA and suppress stale follow-up

Confidence81%

Setup and provider boundary

Privacy-first demo with a model-pluggable architecture

The public Vercel judge demo works without Gmail login or private keys. For a real workspace, add Google OAuth env values, connect readonly Gmail, add a server-side OLLAMA_API_KEY for Ollama Cloud, then sync recruiting mail into the same review-gated agent pipeline. Other hosted providers remain adapter or BYOK extension points unless explicitly implemented.

01.env.localSet Google OAuth and optional OLLAMA_API_KEY.

02Connect GmailReadonly OAuth stores a local token under .careeros-data.

03Enable Ollama CloudCareerOS calls https://ollama.com/api from the server.

04Review outputMailbox triage, extraction, evidence, and notifications stay inspectable.

Deterministic fallbackAlways availableRules-based parsing. No model, no key, no downloads. Always available and always the fallback path when a model-backed run fails.

Gemma via Ollama CloudPrimary cloud pathPrimary model path. CareerOS calls Ollama Cloud directly at https://ollama.com/api with OLLAMA_API_KEY. Output is schema-validated and review-gated.

MLX (Apple Silicon)Advanced adapter roadmapNative Apple Silicon runtime candidate. Would integrate as a `local-runtime` adapter once a stable JSON/CLI bridge to the user's local MLX models is designed. Need a bounded local-process boundary (subprocess or local HTTP) plus schema-validated output before CareerOS opens a second local runtime.

llama.cppAdvanced adapter roadmapPortable C++ runtime that can host Gemma 4 quantizations. Research-only until CareerOS has a separate advanced-runtime boundary. Same boundary as MLX: stable local endpoint + schema validation + tests.

LiteRTAdvanced adapter roadmapGoogle's on-device runtime (LiteRT, formerly TFLite). Best fit for the mobile-wearable hackathon track; unlikely to be the primary path for a desktop console. Needs a mobile/edge surface story before CareerOS embeds a LiteRT pipeline; out of scope for the laptop-first public RC.

vLLMAdvanced adapter roadmapHigh-throughput GPU serving runtime. Useful for self-hosted Other-Candidate scenarios, not part of the public demo runtime. Needs a deployment surface other than the laptop dev script. Belongs to Other Candidate hosted, not the public RC.

SGLangAdvanced adapter roadmapStructured-generation-friendly serving runtime. Compelling for the JSON-only output shape CareerOS already uses, but heavier than Ollama for first-run users. Same deployment-surface constraint as vLLM.

Gemma 4 MTP draftersAdvanced adapter roadmapMulti-token speculative drafters for lower-latency Gemma inference. Research-only until a stable upstream runtime API exists. Wait for the upstream runtime to publish a stable drafter API and benchmarks before exposing it as any runtime option.

OpenAIBYOK adapter roadmapBYOK hosted adapter. User would provide their own key; CareerOS would only read structured JSON output and still route it through the review gate. Blocked on encrypted local credential storage + redaction rules. Will not ship before that boundary is reviewed.

AnthropicBYOK adapter roadmapSame BYOK shape as OpenAI; same model-router interface; same review-gate boundary. Blocked on encrypted local credential storage + redaction rules.

OpenRouterBYOK adapter roadmapAggregator BYOK route. Lets the user pick a hosted model without CareerOS bundling provider-specific code; same credential boundary as above. Blocked on encrypted local credential storage + redaction rules.

Reproducible case

Assessment evidence reaches a typed proposal and stops at review before tracker mutation.The checked-in fixture, result JSON, and case study document the same boundary.