KUNGUM ABALA JI

Architected and owned a production conversational voice AI platform for customer service and booking, integrating telephony with native audio speech-to-speech. Migrated the conversational path from cascaded STT→LLM→TTS to a persistent LiveKit + Gemini native audio pipeline, using WebSocket sessions, RAG with entity extraction, custom VAD/noise handling, fallbacks, and Salesforce automation. Served 150+ unique customers, automated the journey from service discovery through booking capture, instrumented Salesforce to track conversion and revenue impact, and improved customer experience. Built hybrid maintenance RAG using sentence transformers embeddings with FAISS semantic retrieval plus keyword/metadata signals to ground responses in SOP guidance and support issue categorization/service trade-declassification prior to Salesforce actions. Designed a governed text-to-SQL analytics agent for ~25 senior stakeholders with query rewriting, retrieval/reranking, SQL generation and guardrai

Built a multi-agent IT/admin support platform for ~1,500 internal users, reducing dependency on a two-person support team through query rewriting, classification, vision, knowledge grounding, and Salesforce-connected agents with human-in-the-loop escalation. Engineered ETL/ELT pipelines over roughly one year of noisy WhatsApp support history using NLP-driven cleanup and deduplication, and redesigned retrieval around hierarchical page/section/answer nodes to narrow the search space before LLM generation. Implemented MCP-based Salesforce access and LangSmith tracing/evaluation with a golden dataset for routing/retrieval/answer regression testing. Delivered a mobile PWA and migrated deployment from Cloudflare to GCP with CI/CD and a 90/10 stable/canary release strategy. Designed and implemented a hybrid long-term memory system for agents (vector store with metadata + LLM memory agent distillation to structured JSON, entity-linked vector memory, and a SQLite short-term buffer of last 10 m

Designed a multi-agent treasury AI decisioning system in Python/Crewa AI to assess transaction purpose, fraud/anomaly risk, company policy constraints, approval thresholds, liquidity, and cash reserves before automated processing or human escalation. Built a multi-user receipt and expense-document extraction service that classifies uploaded documents and auto-fills a financial ledger, serving Qwen 3-VL-L4B through a dedicated vLLM OpenAI-compatible GPU service with JSON-schema constrained decoding and Pydantic validation, plus deterministic decimal-based arithmetic checks and a human-labeled benchmark (schema validity, exact-match money accuracy, hallucination rate, and p50/p95 latency) to gate rollout. Developed LSTM and transformer volatility forecasting experiments across multiple asset classes to compare time-series behavior and prediction quality for stakeholder risk and exposure analysis.