Hi, I'm Anjana 👋

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.

🔬 What I'm working on

🧠 Lab Rotation 2 @ HHU — DPO-based hallucination reduction on dialogue NLG. Strict preferences beat gold/random by 4–6× on faithfulness; working paper underway.
🏭 RIF Institut, Dortmund — Containerised brewery process control: PLC over MQTT, time-series replay via FastAPI + InfluxDB, Streamlit ops dashboards.
📈 Recently wrapped — TFT + PPO pairs trader: an end-to-end, CI-tested forecasting + RL stat-arb pipeline whose look-ahead-free walk-forward honestly showed no edge over a SPY buy-and-hold (the value is the rigour, not a backtest Sharpe).

⭐ Featured work


🖥️ wattson	DL-workload-aware system monitor — `textual` TUI with GPU-aware process table, NVML temps, 1 Hz refresh. Active development.
📈 tft-ppo-pairs-trader	Temporal Fusion Transformer spread forecaster + PPO RL agent for multi-asset pairs trading (DuckDB · Gymnasium · Optuna/MLflow · Docker · CI, 41 tests). Built a look-ahead-free walk-forward backtester — hedge ratio and TFT refit per fold — that honestly showed the strategy has no edge over a SPY buy-and-hold: a pipeline rigorous enough to reject its own false positives.
🔧 sensor-anomaly-detection · 🚀 demo	Unsupervised anomaly detection on NASA C-MAPSS turbofans. 5 detectors across FD001–FD004; SHAP attribution surfaced that the model picks up HPC degradation through bypass-duct pressure before direct sensors drift. Streamlit + Power BI surfaces.
🔬 Digital-IC-NN-Duplication	NN-based digital twin for sequential circuits. Published at IEEE 2023.
🖼️ image-downsampling-processor	Verilog low-pass + downsample-by-2 processor, validated against Python reference via SSD.