Oct 2025 – Present · 7 mos · Remote

• Promoted after Outstanding (5/5) Rating
• Built HITL authentication workflow with Slack integration, reducing
• customer response time 75% (20m → 3m) and improving Resolution rate by
• 3‐5%/week through conditional approval flows.
• Expanded knowledge base through Q&A extraction (Memory) from
• recordings, improving retrieval accuracy.

Apr 2025 – May 2025 · 1 mo · Remote

• Mentoring an AI Engineer intern by helping him solve a business challenge involving Request for Proposal (RFP). Helped with setting up OpenAI agents SDK, designing the Agentic solution, setting a learning plan, solved the unstructured excel data extraction issue he was facing etc.

Mar 2025 – Oct 2025 · 7 mos · Remote

• Working as a Generative AI Engineer:
• Caching cut quadratic cost growth in multi-turn chats: a 5-response ticket dropped from $30 → $12 (60% savings). Repeated trace cost fell 99.97% ($3 → $0.001) with 150–450x faster responses (<100ms). Projected annual savings: $21.6K–$43.2K across 2,400 tickets
• Designed and implemented an API connected with Salesforce that captures customer support engineer feedback, updates AI response records in MongoDB, and provides internal teams with structured data for quality monitoring and process improvements
• Implemented intelligent database caching with 7-day refresh logic for web scraping operations, reducing response time from 15-18 seconds to 2-3 seconds (85% performance improvement).
• Shifted the existing project from CrewAI to OpenAI Agents SDK framework in less than a day for POC.
• Implemented CrewAI guardrails to enhance response validation for the support bot, preventing bad UX by validating and rerunning faulty tasks. This reduced customer churn and improved email response quality
• Developed a VWO Campaign Report Tool, overcoming missing documentation by manually mapping different settings to their actual backend IDs.
• Continuously reviewed support tickets addressed by the bot, identified issues, and implemented actionable improvements to enhance its performance.
• Made 100s of prompt updates to refine responses and enhance bot's accuracy.
• Added custom context for various support issues by communicating with the support team, effectively creating a form of documentation, and implementing it into the support bot. It's being used like a custom vector database.
• Took a chat project out of a limbo by providing actionable feedback and making prompt updates which helped it get final approval.
• Streamlined the RAG workflow using embedding similarity with NumPy, resulting in an 80% reduction in processing time.
• Improved efficiency by ~25% by running tools asynchronously in Crew, cutting processing time from 8 to 6 minutes.

Sep 2024 – Feb 2025 · 5 mos · Remote

• Developed an AI Agentic workflow for generating a second-tier response in the email support bot.
• Enhanced response quality of AI agent workflow by applying advanced and structured prompt engineering techniques.
• Improved the accuracy and reduced hallucination drastically by making the tool usage in Agentic workflows extremely targeted and removing the unnecessary OpenAI API calls (1000s of AI API calls reduced to zero), which also led to a drastic decrease in token usage.
• Designed a workflow for A/B testing knowledge base embeddings against Google Search within the Agentic framework.
• Refactored Python code for the support bot to improve response accuracy and efficiency.
• Optimized code by reducing 1000 lines of code to 40 lines.
• Parallelized and reduced the time by 80% of the Google Search Content scraping mechanism in the RAG system
• Proposed the idea to use multiple queries in the RAG system
• Improved spam detection to prevent customer replies from being lost in automated email threads, ensuring genuine messages aren’t misclassified.
• Optimized the Human-in-the-Loop (HITL) prompt to ensure Crew AI agents consistently trigger the waiting mechanism. This prevented skips and guaranteed that real support engineers' Slack responses are correctly relayed back to the Crew AI agent.