Experience

• Built and deployed production RAG pipelines serving 10k+ daily users
• Fine-tuned LLMs on domain-specific datasets using RLHF

The loss function for the PPO fine-tuning process relies on the reward model's scalar output:

This approach significantly improved alignment on the core tasks.

Amy – Autonomous Customer & Affiliates Support Agent

• Architected and deployed Amy, a platform-agnostic AI customer support agent spanning web, WhatsApp, and email, designed from the ground up so the AI core remains fully decoupled from any single platform, enabling new channel integrations without touching business logic; automated 77% of 7.2K+ daily conversations while sustaining 95%+ CSAT across 33K+ daily messages in production.
• Engineered Amy’s three-layer architecture: an API layer normalizing all inbound platform webhooks, an adapter layer abstracting Intercom and HubSpot behind a unified interface, and an AI layer where a triage agent classifies issues and dynamically selects from a library of resolution skills to guide users through fixes with contingency handling for uncooperative flows, while a parallel knowledge base layer handles dynamic FAQ resolution in real time.
• Evolved Amy into a full multi-agent orchestration platform handling complex account, KYC, and payment workflows through secure data pipelines and end-to-end workflow chaining; structured human-in-the-loop escalation ensures edge cases reach the right agent tier within seconds, making Amy the backbone of an AI-first support operation rather than a surface-level chatbot.
• Built Amy with full multi-modal, multi-turn conversational intelligence, users can send text, images, audio, or documents across multiple consecutive messages and Amy resolves all of them in a single coherent response, addressing each query without losing thread; supports reply generation in text, audio, and image-based charts, enabling rich, free-flow support interactions that feel native to each platform rather than constrained by a bot interface.
• Built an AI-powered partner (affiliates) support version of Amy using RAG (Weaviate + OpenAI embeddings) and Google ADK, enabling autonomous, personalized responses on commissions, referral links, and program policies; engineered a multi-platform chat integration across Intercom and HubSpot with a Cross-Platform Escalation (CPE) system, preserving full conversation context during cross-channel handoffs and eliminating redundant affiliate re-explanations.
• Extended Amy with proactive outbound capabilities that qualify leads and initiate targeted lifecycle-state engagement across all affiliate stages; AI-driven outbound campaigns achieved a 35% engagement rate, a 12-point lift over the pre-automation baseline, demonstrating measurable uplift in affiliate activation at scale.
• Designed and implemented a lifecycle-aware goal system that tailors AI responses by affiliate stage (Onboarding, Active, Stalled, Growing, Scaling, At Risk, Dormant, Churned) using real-time state, improving proactive engagement through stage-specific guidance injected into agent session context.

Nexus – AI Data Analyst for Business Development

• Built Nexus, an AI-powered natural-language analytics agent for Deriv’s Global Partnership team using FastAPI, Google ADK, and OpenAI; enabled non-technical business development executives to query the entire affiliate dataset across Slack, replacing manual analyst workflows with an autonomous agent handling end-to-end question-to-insight pipelines, fundamentally redefining how the BD team operated.
• Designed a hybrid multi-layer knowledge retrieval system combining PostgreSQL, Weaviate vector search, and a temporal knowledge graph (Graphiti) to ground agent responses in structured data, semantic context, and evolving business rules; the system dynamically selects retrieval strategies per query, achieving high answer accuracy across complex multi-dimensional partner datasets.
• Developed LiveBoard, a custom interactive analytics dashboard (React 19, TypeScript, Vite, Tailwind) with real-time conversational refinement, where Nexus generates Plotly and Matplotlib visualizations stored in Supabase Storage; executives could self-serve data exploration, export PDF/HTML reports, and review full session history, collapsing insight-to-action latency from days to minutes.

Platform Intelligence – Tooling, Moderation, and Internal AI

• Architected a scalable tagging infrastructure with multi-dimensional classification (intent, sentiment, resolution, urgency) auto-tagging 7.2K+ daily conversations using 1,140+ dynamic tags across 18 products; designed a visual tree-based tag management UI (React + D3.js) enabling teams to navigate and evolve complex hierarchies without engineering involvement.
• Built queue-based monitoring system with volume threshold alerts and real-time critical case detection (compliance triggers, payment failures, VIP escalations); implemented tier-based Slack notification workflows routing urgent issues to respective teams within seconds, enhancing triage efficiency for 218 agents.
• Deployed SnapTranslate, a Chrome extension for real-time web translation handling 67K+ monthly executions at $3/month, generating$6K/month in savings for 218 CS agents by eliminating third-party translation costs at scale.
• Developed a real-time P2P chat moderation service with ML-based scam detection, reducing monthly reported scams by 83%; built a custom analytics dashboard enabling Customer Support and Risk teams to conduct proactive monitoring and incident resolution.
• Built Amber, an agentic RAG-powered HR Slack bot serving 1,500+ employees with location-specific policy queries, org hierarchy navigation, and personal HR data access; integrated a daily pulse survey system with department-level dashboards for real-time cultural health tracking.

• Architected scalable data pipelines processing over 5TB of training data daily.
• Optimized model inference latency by 40% using TensorRT.

Below is a simplified architecture of our inference pipeline:

graph LR
  A[API Gateway] --> B[Load Balancer]
  B --> C[Model Server 1]
  B --> D[Model Server 2]
  C --> E[(Feature Store)]
  D --> E

• Prototyped a web dashboard for visualizing model predictions in real-time.
• Wrote data cleaning scripts in Python that reduced preprocessing time by 20%.