ChromaDB · BM25 · Reciprocal Rank Fusion · Streamlit · Gemini / Groq

• Hybrid retrieval combining dense embeddings with a custom Ethiopic BM25 via Reciprocal Rank Fusion: Hit@1 77.51% and Hit@3 92.10% on a 329-question holdout (dense-only Hit@1: 72.64%).
• Bilingual prompt-injection defense, grounded refusal on out-of-scope questions, and an honestly documented re-ranker investigation (tested, too slow on CPU, left disabled).
• Live demo with streaming answers and inline source citations; 29 tests and CI.

Live Demo → · Repository →

PyTorch Geometric · GCN · XGBoost · NetworkX · KS Drift Test

• Benchmark on the Elliptic Bitcoin dataset (203k transactions, 234k edges, 49 time steps) with a strict temporal split.
• Compared tabular XGBoost, hand-built graph features, and a 2-layer GCN.
• Diagnosed a collapse at time step 43 as distribution shift (KS test p < 10⁻⁶⁵), with model confidence falling from 0.933 to 0.040.

Explore Repository →

PyTorch · FAISS · FastAPI · Streamlit · Docker

• Two-tower retrieval model on MovieLens 100K with a per-user temporal split: Hit@10 of 56.6% vs 27.7% for a from-scratch matrix factorization baseline and 39.3% for popularity, under the same protocol.
• Documented the six bugs fixed on the way from a 2.54% first run, including loss bias, false negatives, vocabulary leakage, and an unfair baseline.
• FAISS index served through FastAPI with a Streamlit demo; tests, Docker, and CI.

Explore Repository →

MLflow · Optuna · Evidently AI · FastAPI · Docker

• Risk API comparing a from-scratch Gaussian anomaly model against a supervised baseline, with honest benchmarking (PR-AUC, ROC-AUC).
• MLflow experiment tracking, Optuna hyperparameter tuning, and Evidently AI drift monitoring.
• Containerized with Docker and covered by 18 automated tests.

Explore Repository →

LangGraph · ReAct Agent · Pydantic v2 · SQLite

• Cyclic LangGraph ReAct agent that investigates alerts with tool calls across KYC, transaction-velocity, and sanctions data.
• Dual-mode design: live LLM reasoning (Gemini / Groq) with a deterministic fallback scoring engine that needs no API.
• Pydantic-validated verdicts with auditable JSON execution traces; 23/23 tests passing.

Explore Repository →

TensorFlow / Keras · CNN · NumPy · scikit-learn

• CNN reaching 95.8% test accuracy on the 33 base Fidel characters (3-seed range 95.4–96.5%), vs 86.3% for a dense network and 82.4% for logistic regression.
• Forward pass first written by hand in NumPy; error analysis traced the main confusion (ደ → ጸ) through a manual review of 342 images, and two class-imbalance fixes were tested.
• Extended to the full 238-class syllabary: 86.86% (single seed).

Explore Repository →

Selenium · CatBoost · scikit-learn · Pandas · 5-fold CV

• Scraped 5,046 listings across 10 sub-cities (99.98% success) with a resumable two-stage Selenium pipeline, then cleaned them with per-sub-city outlier detection.
• CatBoost model with R² of 0.826 ± 0.006 (5-fold CV, log scale) and MAE of about 40,800 ETB.
• Per-location results and limitations documented openly.

Explore Repository →

FastAPI · NumPy · Pillow · Docker · pytest

• K-means implemented from scratch (vectorized, multi-restart) to extract dominant color palettes from uploaded images.
• Also returns a web-optimized compressed image (73% size reduction on a sample photo).
• 24 tests, including regression tests for bugs found along the way; Dockerized.

Explore Repository →