How Orin works
A technical deep dive into the intelligence model, the overnight pipeline, and how scores are produced safely.
The intelligence model
Orin maintains a per-user model rebuilt every 24 hours and cached in Redis. It draws on profile data, behavioural signals, network activity, and traction inputs. Because the model is cached, most personalised surfaces render instantly without re-computing.
The overnight pipeline
Scheduled jobs run on Celery beat (Asia/Kolkata) on a singleton worker. Each job is idempotent — guarded by a Redis lock — so a retry never double-sends an email or double-queues an action. The pipeline rebuilds models, scans patterns, rebuilds the knowledge graph, drafts intros, and composes briefs.
Hybrid scoring
- —Deterministic — profile completeness and pattern matching. No LLM cost, fully predictable, used for fast scores like profile strength.
- —LLM — Azure OpenAI
gpt-5.4, temperature 0.1 for scoring and 0.3 for recommendations, with JSON-enforced output. Used for nuanced fit (co-founder, investor, job, mentor).
Resilience & fallbacks
- —Retry with backoff on transient AI failures.
- —A circuit breaker stops calling the provider after repeated failures and auto-recovers within ~60 seconds.
- —Cached scores are served when the provider is unavailable, so the UI never blocks.
- —A platform-wide daily AI budget guards against runaway cost.
Prompt-injection defences
User-supplied text is treated as data, never as instructions. Outputs are constrained to JSON schemas and numeric ranges, and Orin's scoring weights and prompts are never modified by user input.