I build intelligent systems that bridge hardware and software. Currently focused on reducing hallucinations in task-oriented dialogue via preference-based optimisation, industrial IoT automation, and applied ML for time-series problems.

Currently available for new roles: Master's thesis, internships, or workstudent positions in ML/DL. Based in Germany, open to relocation.

• 🕸️ hn-knowledge-graph: Public-facing knowledge graph of the HN AI/ML landscape. Typed NER via local llama3.1:8b (Ollama + LangChain) over ~240 stories, alias normalisation, NetworkX co-mention graph (172 entity nodes / 244 edges after pruning), Louvain communities (top-8 coloured), cytoscape.js viewer at anjanamb.github.io/hn-knowledge-graph. Click any entity to jump to its source HN threads.
• 🤖 hn-ml-trends: LangGraph agent for Hacker News AI/ML trend tracking. Silver labels via qwen2.5:14b (Ollama) + Claude spot-check for the ceiling estimate; TF-IDF baseline 43.3% vs llama3.1:8b classifier 62.7% test accuracy; sentence-transformer clustering; conditional drill-down node fires on real category spikes (research +13.3 pp this window). All eight notebooks executed with baked-in outputs.
• ⚡ electricity-demand-kalman: Classical state-space walkthrough on PJME hourly load. Hand-written Kalman filter + Ornstein-Uhlenbeck residual fit + LightGBM overlay, stacked over a per-month additive decomposition. Full-stack MAPE 0.59% vs persistence 3.27% on the held-out 2016-Q4 test window.
• 🧠 Lab Rotation 2 @ HHU (submitted August 2026): DPO for hallucination reduction on task-oriented dialogue NLG. LR1-labelled preferences beat gold/random baselines by 13-24x on faithfulness at matched hyperparameters. E2E transfer arm cuts DK-NLI unfaithfulness by 45% relative on top of a reproduced DataTuner SFT.
• 🔧 sensor-anomaly-detection: Onset-of-degradation detection on NASA C-MAPSS FD004. Isolation Forest + DBSCAN over 20 hypothesis-driven features across seven executed notebooks; 99.5% cycle-level agreement between the two detectors.
• 🏭 RIF Institut, Dortmund (Jul 2025 - Jun 2026): Containerised brewery process control. PLC over MQTT, time-series replay via FastAPI + InfluxDB, Streamlit ops dashboards.
• 🎓 Student Teaching Assistant, HHU Düsseldorf (Oct 2025 - Apr 2026): Evaluated and graded deep-learning course assignments; supported exam preparation, proctoring, and student guidance.

ML / Deep Learning

Languages

Data & Infrastructure

IoT · Embedded · Networking

Cloud · Tools

_{💡 Curious devs: poke around the portfolio, toggle ▮ terminal mode and try help. There are easter eggs.}