Utkarsh Chandel

About

LLM workflow/agentic systems practitioner who has helped customers harden an LLM-based incident triage prototype into a trusted daily-use production system by adding observability, audits, confidence gating, and deterministic fallbacks. Brings an SRE-style approach to real-time debugging (trace replay, rollback/canary, safe toggles) and is experienced running developer-centric demos/workshops and partnering with sales on technical qualification and security/architecture artifacts.

Experience

Education

Key Strengths

• Turns fragile LLM prototypes into reliable production systems via input stabilization, task decomposition, and guardrails
• Designs explicit failure modes (timeouts, low confidence, missing data) to prevent model bluffing and ensure safe degradation
• Strong LLM/RAG troubleshooting in real time using trace-based, replayable debugging and clear failure-mode classification
• Applies SRE-style incident response: stop-the-bleeding toggles, targeted fixes, regression replay, and canary rollout
• Builds user trust through transparency (evidence links, cited signals, clear rationale) and bounded automation
• Developer-focused demos/workshops using real failure scenarios, interactive diagnosis, and live break/fix demonstrations
• Effective pre-sales partnership: translates business needs into technical success criteria and produces validation artifacts (architecture, security, guardrails)
• Designed and evolved a security analytics/incident triage backend handling 10–20x ingestion spikes without data loss
• Low-latency investigation views (targeting ~2–5s) via optimized caching and async processing
• Systematic performance debugging using per-stage timers and dependency/lookup analysis
• Mitigated cache stampedes and dependency saturation with tiered caching and singleflight/coalescing
• Built resilient API platforms with strong reliability patterns (timeouts, retries w/ jitter, circuit breakers, backpressure, idempotency)
• Led monolith-to-event-driven migration using shadow traffic, canaries, dual-run, and rollback controls with SLO-based promotion gates
• Proactively identified failure modes (duplicates, replays, out-of-order events, poison messages) and implemented DLQ/quarantine + idempotency to prevent cascading failures
• Implemented production-grade authN/authZ including OAuth2/JWT, RBAC/scopes, and tenant RLS with deny-by-default policies