Uses / Skills
Skills — Every Claim, Cited
Skills are grouped by how well I can back them. Strong skills are used in a system I built or shipped; the qualified group is coursework or offline research and says so. Fortinet figures are my own reported numbers.
🤖 Agentic AI & LLM Systems
| Skill | Where I used it |
|---|---|
| LangGraph | CAD repair loop (CORE Lab research) · ATHENA Production V1 14-node staged pipeline · SHASTRA orchestration |
| LangChain | Fortinet agentic RAG diagnostics (production) · CAD RAG module |
| Agentic RAG | Fortinet — LLM plans multi-step tool calls over telemetry and verifies hypotheses with structured function calling; hackathon prototype to production, ~70% faster mean resolution (self-reported) |
| Tool / function calling | Fortinet (structured function calling) · CAD execution feedback · ATHENA research tools |
| Multi-agent orchestration | ATHENA (team) — deterministic staged workflow with typed handoffs · SHASTRA (team) |
| Structured generation (Pydantic) | ATHENA typed artifact contracts · ARTEMIS serving contract · CAD |
| LLM / VLM-as-judge | CAD visual judgment and ATHENA scene evaluation, used as advisory signals, not correctness oracles |
Stack: LangGraph, LangChain, Pydantic
👁️ Vision-Language Models & Retrieval
| Skill | Where I used it |
|---|---|
| CLIP / SBERT encoders | CERBERUS — frozen CLIP ViT-L/14 and MPNet alignment |
| Cross-modal retrieval | CERBERUS — R@5 77.7% / 77.5% (image-to-text / text-to-image) at 4096-d on PixMo validation |
| Matryoshka embeddings | CERBERUS — nested prefixes down to 128-d (R@5 62.9% / 64.0%) |
| LoRA fine-tuning | CERBERUS only — Qwen-2.5 7B, r=64 on q/k/v/o, image-QA SFT; training loss only, no held-out QA metric yet |
| VLM routing & evaluation | ARTEMIS — offline research; leakage-corrected 67.7K-question benchmark, oracle 65.4% vs 49.1% best single model, learned router did not beat lookup (negative result) |
Stack: PyTorch, Hugging Face Transformers, CLIP, SBERT, PEFT
⚙️ Production ML & Inference Systems
| Skill | Where I used it |
|---|---|
| Inference optimization (ONNX Runtime) | Fortinet — quantized CPU-only edge inference on access points and switches, ~40% latency reduction (self-reported) |
| OpenSearch at scale | Fortinet — ingestion re-architecture (async I/O + Go), under 50 to 2,000+ events/sec |
| SLA forecasting | Fortinet — 60+ classifiers, 4 categories, automated retraining, 7-day horizon |
| Anomaly detection | Fortinet — DBSCAN on SD-WAN telemetry · patented unsupervised wireless thresholding |
| Backend ML integration (Python/Go) | Fortinet — data-science models into Go and Python services, CPU-only |
| Serving APIs | ATHENA — FastAPI with SSE streaming |
Stack: Python, Go, ONNX Runtime, OpenSearch, FastAPI, Docker, Redis, PostgreSQL, Azure
🔬 Evaluation & Research Rigor
| Skill | Where I used it |
|---|---|
| Automated evaluation harnesses | CAD — ~40-metric STL evaluation stack, pytest harness, paired bootstrap confidence intervals |
| Geometric verification | CAD — voxel IoU, Chamfer and Hausdorff distances on STL meshes |
| Ablations and negative results | CERBERUS Perceiver ablation · CAD feedback matrix and failed pre-registered gate · ARTEMIS router vs lookup |
| Benchmark construction | ARTEMIS — leakage audit that voided 3,909 cross-split image hashes |
| Experiment tracking | Weights & Biases, TensorBoard |
Stack: pytest, Weights & Biases
🛡️ Security & Memory Forensics
| Skill | Where I used it |
|---|---|
| Memory forensics (Volatility 2) | Malware_Analysis — plugin orchestration for feature extraction |
| ML for security | Malware_Analysis — nine scikit-learn classifiers over 976 memory windows (best 88.74%, exploratory); paper accepted at IEEE ICAIA 2026 · Neural-Malware-Shield GraphSAGE on MalNet-tiny |
| Adversarial ML (coursework) | AI-Security — PGD, embedding poisoning, membership inference |
| Dataset curation | Hugging Face Hub — 33 GB analysis outputs plus ~470 GB raw memory dumps |
| Security auditing | Audit_Script_Development |
Stack: Volatility, YARA (vendored rules), scikit-learn, Python
🐍 Core Engineering
| Skill | Where I used it |
|---|---|
| Python | Every system above · 4.5 years in production at Fortinet |
| Go | Fortinet — OpenSearch scaling, backend services |
| PyTorch | CERBERUS training · ARTEMIS router · RL Soccer · Neural-Malware-Shield |
| scikit-learn | Fortinet classifiers · Malware_Analysis |
| SQL / PostgreSQL | ARTEMIS profile store · independent financial-analytics project |
Stack: Python, Go, PyTorch, SQL/PostgreSQL, Linux, Git, Bash
🎓 Academic (explicitly qualified)
| Skill | Where I used it |
|---|---|
| PPO / DQN, curriculum learning (Ray/RLlib) | RL_Soccer_project — academic team project, small-N evaluation |
| Distributed training (DDP, mixed precision) | Coursework code paths only; no multi-node LLM training |
| Post-training literature (RLHF, DPO family, reward models) | Coursework and self-study reading; no hands-on post-training of my own |
💻 Daily Drivers
- Compute: Georgia Tech HICE GPU cluster
- Editor / Terminal: VS Code · Zsh + tmux
- Infra: Docker, Linux (daily driver), Git, Azure