Velric

Postgres-as-queue admission control, the generation_jobs table IS the queue; jobs are claimed with an atomic optimistic lock so any web instance or the dispatcher can fire safely, capped by a global concurrency limit.

Two-tier backpressure with load-shedding, a fair dispatcher plus a hard VM-side semaphore, a max-queue-depth that sheds load with 503s, and a stale-slot sweeper that reclaims jobs stuck past a 10-minute timeout.

Opportunistic + autonomous drain, dispatches inline on new requests, and a 10-second cron sweep drains the backlog FIFO when traffic goes quiet.

Background generation UX, kicks generation off on dashboard mount to hide a 30s+ blocking wait that previously caused production 504s.

AI code-sabotage engine, GPT-4o injects targeted bugs into blueprint code and personalizes the story to the candidate, running personalization and sabotage concurrently.

Closed-loop sabotage validation, executes the sabotaged code against the blueprint's tests to confirm the bug actually breaks the suite, retrying and cross-checking the produced error before it's ever served.

Warm mission pool, matches pre-generated missions by type, difficulty and category with Postgres array-overlap, clones one per candidate for an instant serve, and retires a mission after a set number of serves.

Difficulty-adaptive tuning, adjusts time limits and hint verbosity from the gap between blueprint difficulty and the candidate's stated experience.

pgTAP SQL grading in throwaway databases, spins a uniquely-named temp Postgres DB per run, installs pgTAP, runs the candidate's setup and test SQL, parses the TAP output, and drops the database in a finally cleanup.

Multi-language execution, JS/TS on server-side Jest and Python on pytest, each in its own per-run temp workspace, normalized into one result shape with per-step timeouts and output caps.

AI code-review grader on Claude, scores submissions across code quality, efficiency, security and best-practices, returning a letter grade, strengths, improvements and a feedback paragraph.

Honest, process-level isolation, ephemeral databases, temp workspaces, OS process boundaries and hard timeouts, torn down after every run.

Seven-queue BullMQ topology, matching, assessment, grading, tech-grading, candidate-summary, job-description and doc-verify, each with dedicated worker processes.

Resilient job defaults, three attempts with exponential backoff, completed jobs retained a day and failed jobs a week for debugging, all Sentry-instrumented.

PM2 cluster orchestration, runs the web app at six cluster instances (deliberately below the box's core count for headroom) plus dedicated processes per AI worker and the generation dispatcher.

Self-hosted ops, a two-VM split (prod app and workers vs the agentic FastAPI service), both managed through PM2.

Vision-based DocVerify, extracts name, institution and GPA from PDFs and images with GPT-4o vision, runs an authenticity check for tampering, and does soft fuzzy name-matching plus duplicate-profile detection.

Résumé-parsing pipeline, PDF/DOCX extraction → GPT-4o structuring → schema validation with retries → sanitized JSON safe for Postgres jsonb.

Anti-bluffing rubric grader, GPT-4o grades submissions across five weighted dimensions, then generates verification questions from the submission and re-grades the answers to adjust the tier.

Behavioral anti-cheat, folds tab-switch counts, attempts and time into the final score with explicit penalty deductions.