I build production-grade LLM-powered agents, multi-agent pipelines, and RAG systems — deployed on cloud infrastructure, evaluated rigorously, and built to hold up under real conditions.

Four years of deliberate investment in applied AI: freelance production work, structured academic programmes (MIT-IDSS, HarvardX, DataCamp), and competitive hackathon builds — bridging enterprise engineering discipline from Germany with frontier AI research.

M.Sc. thesis on large-scale geospatial data processing adopted into live production at FusionSystems GmbH for the Galileo Map Service.

Agentic AI — Multi-agent orchestration · Tool-use · ReAct pattern · Stateful agents · LangChain · Conversational AI
LLMs & GenAI — OpenAI API · Gemini 2.5 Flash · Prompt engineering · LLMOps · Evaluation design · Fine-tuning concepts
RAG & Retrieval — Semantic search · Vector DBs (Pinecone, FAISS) · Embeddings · Chunking · Hallucination grounding
MLOps & Cloud — GCP · Vertex AI · Cloud Run · Docker · Kubernetes · Cloud Build CI/CD · Datadog · MLflow · AWS · Azure
Languages — Python · C# · ASP.NET Core · Node.js · FastAPI · SQL · JavaScript · PHP · Bash
Data & Databases — MS SQL Server · MongoDB · ETL pipelines · XML schema design · Geospatial data processing

Google Gen AI Academy APAC · Track 1 · 2026

Full-stack production multi-agent system built end-to-end during the Google Gen AI Academy APAC programme.

• Agentic pipeline: React + Vite → Node.js Express orchestration → Gemini 2.5 Flash reasoning layer → persistent storage
• Multi-agent orchestration: Invoice creation, payment tracking, and real-time financial analytics via stateful tool calls and session memory
• LLM reliability engineering: Diagnosed and resolved a silent model/API version mismatch causing production failures — no crash, just wrong outputs. Traced to orchestration layer, resolved under live load
• Deployment: Containerised with Docker · Google Cloud Run · Cloud Build CI/CD
• Impact: ~70% faster invoice workflows

Scalar Hackathon · 2026

Agentic workflow engine that interprets natural language intent and autonomously orchestrates multi-step workflows across tools and environments.

• NL-to-action: Agent decomposes free-form user instructions into structured multi-step tool call sequences
• Robust execution: Maintains environment state · detects tool failures · classifies failure type · retries with corrected parameters
• Modular design: Tool registry allows new integrations without modifying core agent logic
• Reliability pattern: Production-grade recovery logic — the hardest part isn't the happy path, it's knowing when to retry vs when to stop

DEVPOST Hackathon · 2025

Multimodal agentic system for automated root-cause analysis of production incidents.

• Integrated Datadog metrics, logs, and dashboard screenshots with Vertex AI (Gemini) for AI-driven incident diagnosis
• Designed evaluation metrics to measure AI diagnostic accuracy against ground-truth incidents
• FastAPI service deployed on Cloud Run (serverless)
• Structured AI insights significantly reduced incident investigation time

Personal Research Project · 2023–25

LLM-powered semantic retrieval agent with prompt chaining and contextual memory.

• Embeddings + vector DB for semantic retrieval
• ~40% faster resolution vs keyword search
• Containerised via Docker/CI-CD — production-ready pattern

M.Sc. Thesis → Live Production Deployment · TU Chemnitz & FusionSystems GmbH · 2016–17

Research combining OpenStreetMap and NASA SRTM satellite data via spatial interpolation (IDW) to build a queryable 3D geospatial database — adopted into live production at FusionSystems GmbH for the Galileo Map Service.

• Pipeline: OSM PBF + SRTM DEM ingestion → PostgreSQL/PostGIS → IDW spatial interpolation → 3D geometry (LineStringZ, PolygonZ) → GeoServer rendering with custom SLD map tiles
• Full ML lifecycle: data processing → modelling → deployment → monitoring at scale

Freelance AI/ML Engineer & Researcher — Independent · Apr 2022–Apr 2026
AI/ML and data engineering solutions for European clients · RAG systems · Semantic search agents · Multi-agent pipelines

Full Stack Software Developer — PICA GmbH · Munich, Germany · 2018–2021
Enterprise software for BMW (SAP invoice automation · 80% accuracy improvement), Otto Bock (orthopaedic caretaker app), and the Bavarian State Medical Association (nursing training platform)

Software Developer (Contract) — AMAN Media GmbH / AAM IT GmbH · Munich · 2018
Thermomix backend platform for Vorwerk Switzerland · large-scale DB migration · ETL automation · anomaly detection

M.Sc. Information & Communication Systems — TU Chemnitz, Germany · 2013–2017
B.E. Electronics & Communication Engineering — Visvesvaraya Technological University, India · 2009–2013

• Reliability over hype — production systems fail in ways demos never do
• Evaluation-first — if you cannot measure it, you cannot improve it
• Research to production — every project runs the full lifecycle
• Fail fast, recover smarter — the recovery logic is the hardest and most important part

📍 Bangalore, India · Open to relocation
📫 darshanl1711@gmail.com
🔗 LinkedIn