Uses / Skills
Skills โ Every Claim, Cited
Every skill below is backed by a real system I built. Evidence links point to my code, repos, or the specific project where it was used. Nothing on this page is aspirational.
๐ค Agentic AI & LLM Systems
| Skill | Where I used it |
|---|---|
| LangGraph (repair loops, supervisor-worker, planners) | CAD pipeline repair loop (CORE Lab research) ยท ATHENA supervisor-worker routing ยท SHASTRA planner |
| LangChain | Fortinet agentic RAG diagnostics (production) ยท CAD RAG module |
| Agentic RAG | Fortinet agentic RAG diagnostics โ LLM plans multi-step tool calls over telemetry, verifies hypotheses via structured function calling; hackathon prototype โ production, ~70% mean resolution-time reduction |
| Tool / function calling | Fortinet (structured function calling) ยท CAD execution feedback ยท ATHENA tools.py |
| Multi-agent orchestration | ATHENA โ 5 specialized agents + supervisor-worker via LangGraph Command, reflection critique |
| Structured generation (Pydantic) | ATHENA (OrchestrationRouter / ReflectionRouter) ยท SHASTRA ยท CAD |
| LLM-as-a-judge | CAD VLM-based visual judgment ยท ATHENA reflection loops |
| Planning & reasoning | SHASTRA trace-to-graph ยท CAD multi-step repair ยท ATHENA |
Stack: LangGraph, LangChain, Pydantic
๐๏ธ Vision-Language Models & Retrieval
| Skill | Where I used it |
|---|---|
| vLLM serving | ARTEMIS โ 5 VLMs across 10 endpoints (Gemma 3 27B, Qwen3-VL, Qwen2.5-VL, DeepSeek OCR) ยท CERBERUS |
| VLM routing & evaluation | ARTEMIS โ trained neural multi-task router, SLA-aware load balancing (simulation-validated), 5 routing modes, ~340K profiles / ~68K queries, 90.3% oracle-utility recovery (balanced) |
| CLIP / SBERT / FAISS | CERBERUS โ frozen-encoder alignment + retrieval |
| Cross-modal retrieval | CERBERUS โ R@5 ~78% on PixMo; 4096โ128-dim Matryoshka compression |
| Matryoshka Representation Learning | CERBERUS โ src/encoders/mrl.py, prefix-sliced projector |
| LoRA fine-tuning of Qwen2.5 decoders | CERBERUS โ PEFT r=32 ฮฑ=64, Qwen2.5-7B/3B/1.5B on GT HICE cluster |
Stack: PyTorch, Hugging Face Transformers, vLLM, CLIP, SBERT, FAISS, PEFT
โ๏ธ Production ML & Inference Systems
| Skill | Where I used it |
|---|---|
| Inference optimization (ONNX Runtime) | Fortinet โ edge deployment on network appliances, ~40% latency reduction |
| OpenSearch / Elasticsearch at scale | Fortinet โ full ingestion re-architecture (async I/O + Golang), 50 โ 2,000 events/sec (40x) |
| SLA forecasting | Fortinet โ 60+ classifiers, 4 categories, automated retraining, 7-day horizon |
| Anomaly detection | Fortinet โ DBSCAN on SD-WAN telemetry (reportedly prevented >50% of potential outages) ยท unsupervised wireless thresholding (patent) |
| Backend ML integration (Python/Go) | Fortinet โ distributed telemetry systems, CPU-only pickle model serving |
| ML serving APIs | ARTEMIS โ FastAPI inference stack over Postgres-backed profiles |
Stack: Python, Go, ONNX Runtime, OpenSearch, FastAPI, Docker, Redis, PostgreSQL, Azure
๐ฌ Evaluation & Research Rigor
| Skill | Where I used it |
|---|---|
| Automated evaluation harnesses | CAD โ pytest harness, compile + geometric gates on every run |
| Geometric verification | CAD โ Chamfer/Hausdorff on STL meshes, precision/recall/F1, normal consistency |
| Ablation study design | CERBERUS โ Perceiver Resampler ablation (honest negative result) ยท CAD repair-loop components |
| Cross-modal evaluation | ATHENA โ BLEU/ROUGE-L/METEOR/BERTScore + CLIPScore/SSIM/PSNR on 100 reference videos |
| Experiment tracking | CAD + ARTEMIS + CERBERUS โ Weights & Biases, TensorBoard |
| Benchmark construction | ARTEMIS โ 5 evaluation suites (VQA, OCR, captioning, reasoning) ยท CAD on CADPrompt benchmark |
Stack: pytest, Weights & Biases, TensorBoard
๐ก๏ธ Security & Memory Forensics
| Skill | Where I used it |
|---|---|
| Memory forensics (Volatility3) | Malware_Analysis โ automated orchestration of malfind, pslist, vadinfo, yarascan |
| YARA rule development | Malware_Analysis โ 100+ rules for ransomware family classification |
| ML for security | Malware_Analysis โ scikit-learn pipeline for malicious-process identification; published at IEEE ICAIA 2026 |
| Adversarial robustness | AI-Security โ PGD attacks, embedding poisoning, blind backdoors, model extraction, membership inference, watermarking (coursework) |
| Dataset curation at scale | Hugging Face Hub โ 33 GB analysis dataset + ~470 GB raw dumps + 3,384 code files |
| Security auditing | Audit_Script_Development โ automated Linux posture-audit tooling |
Stack: Volatility3, YARA, scikit-learn, Python
๐ Core Engineering
| Skill | Where I used it |
|---|---|
| Python (expert) | Every system above ยท 4.5 years production at Fortinet |
| Go | Fortinet โ OpenSearch scaling, backend services |
| PyTorch (DDP, mixed precision) | CERBERUS โ distributed training on H100/A100 ยท ARTEMIS router training |
| Scikit-learn | Fortinet 60+ classifiers ยท Malware_Analysis pipeline |
| Distributed training | CERBERUS โ PyTorch DDP + mixed precision, GT HICE cluster (H100/A100) |
| HPC / Slurm | CAD multi-GPU runs ยท ARTEMIS/CERBERUS cluster jobs |
Stack: Python, Go, PyTorch, SQL/PostgreSQL, Linux, Git, Slurm, Bash, C/C++ (coursework)
๐ Academic (explicitly qualified)
| Skill | Where I used it |
|---|---|
| PPO / DQN, reward shaping, curriculum learning, self-play (Ray/RLlib) | RL_Soccer_project โ 2v2 soccer agents (academic project, team) |
| Post-training literature (RLHF/DPO-family, process-vs-outcome reward models) | Coursework/self-study notes tied to CAD verifiable-reward design โ literature familiarity only, no hands-on production post-training |
๐ป Daily Drivers
- Compute: Georgia Tech HPC (Slurm), H100/A100 clusters
- Editor / Terminal: VS Code ยท Zsh + tmux
- Infra: Docker, Linux (daily driver), Git, Azure