Daniel – Python, LLM, RAG

An autonomous conversational sales agent deployed over WhatsApp to qualify and close cold outbound leads for a public-servant loan product. The agent handles multi-product flows, simulates loan terms in-process, and hands off to human agents at the documentation stage. Achieved ~2× the human agent conversion rate (~5% → ~10%) in production.

• Designed and deployed a production LangGraph-based conversational agent over a WhatsApp → RabbitMQ → AWS pipeline, with phone-number-keyed state checkpointing for stateful multi-turn conversations.
• Implemented an intent router with hardcoded FAQ responses for high-frequency queries, avoiding LLM cost and latency on common turns and reducing per-conversation token spend.
• Built in-process loan simulation tools to bypass a timeout-hostile async vendor API, enabling real-time product calculations mid-conversation.
• Doubled cold-lead conversion rate from ~5% (human agents) to ~10% (AI agent) as the sole developer; instrumented funnel events to surface the delta to stakeholders.
• Built a mixed eval harness — deterministic code assertions (e.g., scam-concern guardrail checks) plus LLM-as-judge for structured refusal quality — with CI integration and ~15-run pass-rate statistics per case.
• Integrated Opik for end-to-end LLM tracing; applied prompt-prefix caching discipline (stable system prefix, variables appended) to control costs while validating model value.
• Added Grafana dashboards tracking LLM response time by negotiation stage and error rate via Prometheus and OpenTelemetry.
• Developed a custom Claude Code skill to automate eval generation, programmatically scaling test suites before each deployment.
• Conducted a feasibility study on LLM fine-tuning, evaluating dataset volume requirements and use-case fit against RAG and prompt engineering alternatives.

A conversational AI assistant built on a full-stack Next.js and Django application. The system incorporated RAG and GraphRAG pipelines for document-grounded responses over a hybrid PostgreSQL/Neo4j knowledge store, plus TTS integration for voice output.

• Architected RAG and GraphRAG pipelines using PostgreSQL (pgvector) and Neo4j, enabling both vector-similarity and graph-aware retrieval for document-grounded conversational responses.
• Integrated TTS for voice output, extending the conversational interface beyond text interactions.
• Built and deployed the full-stack application — Django REST backend, Next.js frontend — on AWS as the sole developer, from initial design through production deployment.
• Drove AI feature decisions independently in a greenfield ambiguous environment, moving from research prototypes to integrated production features.

An internal documentation chatbot built on RAG and a semantic search feature integrated into a live production application, developed within a software consultancy serving diverse global clients. Also included an ML regression model and a feasibility study on SLMs for on-device inference.

• Built a semantic search capability for a live production application using OpenAI embeddings and similarity search; prototyped a RAG-based documentation chatbot over internal knowledge bases.
• Conducted a feasibility study on small language models (SLMs) for on-device inference, evaluating latency and quality trade-offs against API-hosted LLMs.
• Delivered a 61% improvement in predictive accuracy over the prior baseline with a machine learning regression model.
• Owned full-stack development across React and Django — from feature implementation through production deployment on AWS and DigitalOcean, with CI via CircleCI.

A cloud-native infrastructure platform and real-time event processing stack built within a large financial institution's innovation unit. Combined IaC-provisioned Kubernetes clusters with stream processing and Complex Event Processing (CEP) pipelines — an early-stage validation of patterns that predated mainstream cloud-native adoption.

• Built real-time stream processing and CEP pipelines with Apache Kafka, Apache Flink, and Apache Spark to validate data-driven business hypotheses at banking scale.
• Architected IaC-based provisioning and CI/CD pipelines using Terraform, Ansible, and Jenkins — adopted ahead of industry standardization.
• Evaluated and integrated Prometheus, Grafana, and Elasticsearch for infrastructure and pipeline observability.
• Turned high-level innovation mandates into working prototypes, bridging strategy and execution in an R&D-focused team.