Four featured builds, then every other repository. Each one links to its code.
Planner, researcher, critic, and synthesis agents draft a cited report, then pause for human approval. Every claim resolves to its source, and approved reports verify offline.
A compiled LangGraph StateGraph with Postgres checkpointing: Planner, then a design gate, Executor and Critic in a loop, Synthesizer, then a human review gate before the report is frozen.
An LLM works out a task in a legacy app with no API once; that run compiles into a typed capability that replays with no model in the loop and hands off to a human when stuck.
Discovery, replay, and human control share one action path: target resolver, policy engine, surface driver, append-only evidence bus. Nothing reaches the UI without an authorized action.
Picks which LLM or pipeline to ship for one job: runs every candidate on your own test cases and ranks them on accuracy, cost, and latency, disqualifying any that miss a hard floor.
Config-driven scoring with a weighted composite and hard constraints, plus a 28-run study comparing single-agent, peer-to-peer, supervisor-worker, and debate-critic topologies.
Plain-English questions over SaaS metrics, answered with streamed tool calls. The model never writes SQL: metrics are declared in YAML, and every tool call is role-checked twice.
One YAML file per metric generates its SQL, argument validation, tool schema, and RBAC scope. Each tool call is authorized at exposure and again at execution, scoped to the JWT tenant.
Research prototypes, coursework built past the brief, and earlier production work.
Engineered an industrial AI assistant for PLC fault log analysis; migrated to Google Cloud (Gemini Flash, Qdrant, Cloud Run, Firebase)
100+ environments · >99% redaction · 30x latency reduction
Engineered a multimodal mental health assessment platform fusing neurophysiological (EEG), vocal biomarker (audio), and behavioral (facial expressions) signals for depression screening. Features Modality Dropout for resilience and interpretability mapped to DSM-5 symptom clusters.
Modality Resilience · HIPAA Aware · DSM-5 Mapping
Built a real-time system recognizing depression by extracting features from facial expressions, EEG signals, and speech modules using multimodal deep learning. Supported by peer-reviewed research publications
3 modalities · 91% precision · IEEE published
Trained YOLOv3 on the COCO dataset for real-time object detection and localization on road scenes; related to autonomous vehicle patents and deployed in a braking system prototype
Real-time detection · 2 patents filed · YOLOv3 + COCO
Developed an ARIMA model built with MATLAB's Econometrics Toolbox to forecast GDP and Stock Indexes of 10 countries. Includes Python-based validation, residual diagnostics, and confidence intervals for economic trend analysis
10 countries · GDP + Stock Index · ARIMA modeling
Architected a data analysis and ML pipeline to find optimal business locations using Wikipedia data and Foursquare API for Toronto neighborhoods; features Folium mapping, K-Means clustering, and geospatial visualization
Toronto neighborhoods · Geospatial clustering · Foursquare API
