ABHIJOY SARKAR

About

Built and shipped PromptGuard, a production middleware proxy that secures GenAI RAG/agent systems against prompt injection and unsafe tool use using risk scoring, graded policy actions, and least-privilege tool gating. Also replaced LangChain abstractions with a custom state-machine runner for a production voice agent to reduce latency and improve traceability, and delivered a clinic call assistant by converting front-desk/doctor requirements into scenario-based guardrails and measurable evals.

Experience

Education

Key Strengths

• Built and deployed production GenAI security middleware to prevent prompt injection and unsafe tool use
• Balances safety and usability via graded enforcement outcomes (allow/redact/warn/confirm) instead of binary blocking
• Strong evaluation discipline: regression suites with normal + adversarial prompts; turns failures into new tests
• Performance-minded engineering: fast-path checks, caching, thresholds, timeouts, and fail-safe defaults to control latency
• High observability approach with structured logs capturing rules/classifiers, risk scores, transformations, and gated tool calls
• Designs reliable agent workflows with explicit state machines, timeouts, retries, fallbacks, and hard tool permissions
• Pragmatic model/retrieval/prompt selection driven by latency/cost/accuracy/safety requirements and A/B testing
• Effective collaboration with non-technical stakeholders by translating needs into scenario checklists, guardrails, and decision-tree flows