Benjamin
Farsijani

Production-grade agentic RAG system that ingests large-scale networking documentation and answers natural-language queries over complex technical content.

• Hybrid retrieval combining dense semantic search (pgvector + text-embedding-3-large) with BM25 sparse retrieval, merged via reciprocal rank fusion; sub-100ms p95 latency over 100k+ embedded chunks
• Multi-step agentic loop with tool use that dynamically re-queries and synthesizes across document chunks before generating a final grounded answer
• Fully containerized ingestion and query pipelines in Docker

Deployed IoT environmental monitoring system for early childcare at the University of Ottawa.

• Wall-mounted ESP32 integrating BME680 (temp, humidity, VOC) and MAX9814 (sound) over I2C; lithium-battery-powered with persistent WiFi streaming
• React Native mobile app for caregivers to monitor real-time and historical classroom readings from any device
• ML anomaly-detection layer flagging air-quality dips, noise spikes, and temperature excursions with real-time alerts

End-to-end algorithmic crypto trading system ingesting live market data across 6 assets, validated via 5-fold walk-forward cross-validation against a buy-and-hold baseline.

• 50+ engineered features: GARCH(1,1) volatility, Parkinson range estimator, VWAP deviation, OBV, ADX, Stochastic RSI, and multi-timeframe log returns with strict zero-lookahead construction
• Probability-gated risk layer: long if p > 0.65, exit if p < 0.35; 3% take-profit / 1.5% stop-loss; 0.1% fee modeling; per-symbol sizing across a $6,000 paper portfolio
• Live containerized dashboard (Flask, Chart.js, Docker, PostgreSQL) tracking equity curve, per-symbol P&L, and trade history against live Binance feeds

Stochastic simulation and Model Predictive Control for steam-assisted heavy oil recovery, targeting a 5-10% reduction in steam-to-oil ratio vs. heuristic baselines.

• Stochastic environment modeling delayed nonlinear dynamics: steam injection to reservoir temperature to production output; 100k+ synthetic multivariate time-series observations
• Constrained MPC with a learned PyTorch state-space model solved in CasADi, optimizing steam injection policies under physical constraints

Palantir Gotham-style threat network intelligence dashboard analyzing actor dynamics during the Iranian Revolution (1979-1981) via GDELT event data.

• Ingested GDELT 1.0 bulk data into a directed weighted NetworkX graph; nodes typed by CAMEO actor codes across Government, Military, Religious, and Opposition factions
• Louvain community detection (20 communities, modularity 0.10); PageRank + HITS composite influence scoring
• Daily z-score anomaly detection surfacing 15 event spikes including the Nov. 4, 1979 US Embassy hostage crisis; interactive Dash + Cytoscape dashboard