01Voice AI — PlatformIn production
Kothon
Vivasoft's AI voice agent platform. I built the agent's memory — a low-latency RAG stack with Docling parsing, Chonkie chunking, Weaviate, and Azure multilingual embeddings — and the telephony connector bridging PSTN calls from PBX with the agent's WebRTC calls via PJSUA/SIP, so the AI talks on real mobile numbers.
PJSUA / SIPWebRTCDoclingChonkieWeaviate
02Sales Ops — AI AgentIn production
Lead Qualifier
Autonomous agent that evaluates every lead from Vivasoft's social media and website. Webhooks pull leads from Pabbly and WordPress, Gemini rates them against qualification criteria, qualified ones flow into the CRM via API, and the sales team gets pinged on Telegram. Hosted on GCP.
GeminiWebhooksPabblyTelegram APIGCP
03Healthcare — RAGIn production
Aurora
Chat system for Health Technology Assessment trial data. CSV ingestion via Azure Functions into Azure SQL, vectorized into Azure AI Search with LangChain + Azure OpenAI. Dynamic filters by country, drug, and agency with full conversation memory.
FastAPILangChainAzure OpenAIAI Search
04Document AI — OCR
Document Extraction Platform
Full-stack document intelligence. Gemini-powered OCR extracts key-value pairs, bounding boxes, and confidence scores; humans validate the low-confidence tail. Template management and role-based workflows, containerized on Azure.
Gemini APIFastAPIMongoDBDocker
05Support — RAG
AiPEN
RAG assistant for internal support and customer queries. Led ingestion, embedding, and GPT response generation. Real-time analytics for query monitoring and feedback trends, plus a feedback loop that keeps sharpening accuracy.
WeaviateAzure OpenAIFastAPIDocker
06HR Tech — AssistantIn production
PiHR Assistant
In-app AI guide for an HR platform — feature help and UI navigation via vector search + GPT-4o mini. Secure isolated backend, and an OpenAI cost-monitoring dashboard so the bill is never a surprise.
GPT-4o miniLangChainWeaviateFastAPI