A production multi-agent research platform that plans before it retrieves. The core insight driving the architecture: open-ended queries against a vector store produce semantically similar results, not relevant results. The fix is structural — agents construct a full report plan first, and every search query targets a specific planned section rather than a vague topic. Retrieval becomes deterministic given a good plan.

• Phase-5 orchestration: 3 manager agents propose full report trees in parallel, a lead agent synthesizes a canonical plan, every query targets a planned section — measurably reducing hallucination and retrieval fanout
• 44-skill auto-registration system (__init_subclass__ metaclass, zero manual wiring) spanning 18 retrieval sources (ArXiv, PubMed, SEC EDGAR, GitHub, ClinicalTrials, YouTube transcripts, legal databases, and more), 18 analysis skills, 8 output skills
• 2-pass credibility-weighted source gate filters low-quality sources before synthesis regardless of semantic similarity score
• Pure BYOK architecture: 10 models across xAI Grok, Gemini 2.5 Pro/Flash, and DeepSeek R1/V3 routed through a single injected LLM client; Fernet-encrypted keys at rest; JWT family rotation with reuse detection
• fastembed/ONNX embeddings — 1.4GB lighter than sentence-transformers, no GPU at inference; deployed as 6-container Docker stack on AWS t3-small at ~$26/month

Stack: FastAPI · ARQ · LangGraph · Qdrant · fastembed/ONNX · Redis · PostgreSQL 16 · Next.js 16 · Docker · AWS · SSE

A career intelligence platform built around the observation that most resume-to-job matching tells you how similar two texts are, not whether you actually meet the requirements. Those are different problems. Wand separates them.

• 6-step wave-parallelized LLM pipeline (profile extraction → JD parsing → company intel → match scoring → contact strategy → action plan) using asyncio.gather(); each step streams its result via WebSocket in real time as it completes
• Multi-model weighted scoring engine: Qualification Match (30%), Technical Skill (25%), Keyword (25%), Formatting (20%) — Gemini 2.5 Pro at temperature=0.0 for nuanced qualification reasoning, Flash for pattern-matching; evidence field enforced non-empty to prevent hallucinated matches
• Unified LLMClient routing all AI calls across Grok, Gemini, and DeepSeek at runtime; all outputs as typed Pydantic models via Instructor — zero JSON parsing anywhere in the codebase
• Multi-source profile unification (PDF resume + LinkedIn export + HTML portfolio) with semantic entity alignment and explicit merge-priority rules

Stack: FastAPI · Celery · Redis · WebSockets · Instructor · Pydantic · Gemini · DeepSeek · SQLite · Docker

• Dual-mode LangGraph system: Chat mode (2–4 LLM calls, 2–4s latency) for quick queries; Think mode (10–15 calls, 15–30s) with Planner → Financial Agent → Publisher workflow and self-correction loops
• 18+ financial tools: quantitative metrics (yfinance), news aggregation (NewsAPI + BeautifulSoup), SEC filing RAG pipeline (10-K/10-Q/8-K → ChromaDB, <200ms vector search)
• MongoDB-backed session persistence; real-time WebSocket streaming; deployed AWS EC2 + S3/CloudFront; 50+ req/sec throughput

Stack: LangGraph · FastAPI · ChromaDB · MongoDB · Gemini 1.5 Flash · React 18 · Docker · AWS

• Fine-tuned LLaVA-1.5/1.6 (7B) with LoRA on InstaCities1M (100k urban images); FastAPI inference server with <2s P90 latency; ~300 req/hr peak
• Full IaC stack: Terraform (VM provisioning on Chameleon Cloud), Ansible (config management), Kubernetes via Kubespray (3-node cluster); MLflow experiment tracking across BLEU, CIDEr, and test loss
• Feedback-to-retraining loop: user corrections → MinIO → Label Studio annotation → MLflow artifact registration; Prometheus/Grafana monitoring

Stack: LLaVA · LoRA · PyTorch · MLflow · FastAPI · Terraform · Ansible · Kubernetes · Prometheus · Grafana

A novel gradient-level intervention for spurious correlation mitigation in CNNs. The key insight: shortcut learning manifests as frequency-dominant gradient energy in a specific band. SGE applies 2D FFT over convolutional kernel spatial dimensions post-backpropagation, decomposes the weight gradient into spurious-band and causal-band components, and attenuates the spurious contribution via a per-channel scaling coefficient — without requiring group labels at training time.

• +3.1pp worst-group accuracy over ERM baseline on Waterbirds benchmark (ResNet-18); +2.5pp post-DFR gain with Tail-loss pairing
• Spurious-to-causal gradient energy ratio exceeds 1.0 from epoch 1 across all runs — shortcut-frequency dominance is an immediate training phenomenon, not late-stage
• Benchmarked 5 objectives (ERM, Group-Balanced, Tail-loss, GroupDRO, JTT); diagnosed GroupDRO failure as calibration stationarity under adversarial reweighting
• Unsupervised calibration via loss-stratified pseudo-groups estimates the radial frequency cutoff — no ground-truth group annotations required

Stack: PyTorch · ResNet-18 · Waterbirds · GroupDRO · JTT · DFR

Lightweight dual-model framework for step-level mathematical reasoning verification. My contribution: the Verifier module end-to-end.

• Fine-tuned TinyLLaMA (1.1B) on PRM800K for step-level correctness classification (correct/unclear/wrong) via 2-phase LoRA — Phase 1 on simple-context examples, Phase 2 on complex multi-step reasoning chains
• 87.91% step classification accuracy; outperformed Phi-2 (81.5%) across QLoRA ablations at rank 16 and 64
• Generator: Phi-2 (2.7B) fine-tuned on GSM8K via QLoRA achieved 64.06% Pass@1 with 7.86M trainable parameters (0.28%)

Stack: QLoRA · LoRA · TinyLLaMA · Phi-2 · PRM800K · GSM8K · HuggingFace PEFT

• Benchmarked EDSR (CNN), SwinIR (Transformer), and Stable Diffusion x4 across DIV2K (natural), TextZoom (scene text), and STAR (astronomical) datasets under zero-shot and domain-adapted protocols
• Introduced Cross-Domain Drop (CDD) metric; diffusion models exhibit catastrophic CDD (79.18% on TextZoom); domain-specific fine-tuning recovers 75–100% of lost performance
• Key finding: text domains require 3x more adaptation (+10.6 dB) than astronomical imagery (+3.4 dB) due to high-frequency structural discontinuities

• FGSM, I-FGSM, PGD, and patch attacks on ResNet-34 under strict l-infinity constraint (epsilon=0.02); PGD reduced Top-1 from 76% to 0.2%
• Confirmed black-box transferability: ResNet-34 adversarial examples degraded DenseNet-121 Top-1 from 74.80% to 59.80% without target model access

• Standardized PyTorch training pipelines (data loading, training loops, evaluation) used by 100+ students across course offerings
• Led debugging sessions on CNN optimization, backpropagation, and autograd — helping students diagnose training failures and performance bottlenecks
• Created supplementary video lectures on ML calculus concepts for a graduate cohort of 50+ students

• Architected distributed microservices (FastAPI + Node.js) serving 1K+ DAU across 10+ clients; maintained 99.9% system availability
• Built full-stack products end-to-end: TypeScript/React/Next.js frontends, ML inference backends, AWS deployments, automated CI/CD pipelines

A 17-minute deep dive into why L1 regularization produces sparsity while L2 shrinks weights — derived from gradient updates and geometric intuition, backed by Python experiments on the Diabetes dataset. Built for students who had memorized the formula but not understood it.