Generative and agentic AI engineer. I started in data engineering, moved to ML platform engineering, and now build LLM systems: agent memory, LLM gateways, fine-tuned models and the evaluations that check them. Each project number below comes from the linked repository. AWS ML Specialty certified.
- 20 pull requests merged upstream, in crewAI, Ray, MLflow, DSPy and 14 other open-source projects (list below).
- Agent memory that beats a hand-tuned rule on held-out data: lethe puts 0.707 of the evidence in the top 10 on LoCoMo vs 0.641 for the best hand-set rule, and only ties it on LongMemEval.
- Security advisory with credit: I reported a sandbox bypass in an AI-agent code executor through a GitHub security advisory; femto/minion adopted the fix and credited the report (GHSA-m5v3-6ccr-cfgx).
Now: doctoral research (DBA, Applied AI) on AI-agent security and LLM robustness. Open to LLM, GenAI and agentic AI engineering roles, internship or full-time.
lethe Go · demo |
0.707 of the evidence in the top 10 on held-out LoCoMo vs 0.641 for the best hand-set rule; on LongMemEval it only ties that rule. A memory server for AI agents that learns from feedback which memories to return, with Python, LangChain and MCP adapters. |
llm-gateway Python FastAPI Redis |
One OpenAI-compatible endpoint over Anthropic, OpenAI and Ollama; 428 tests at 99% line coverage, CI fails under 95%. Response caching, sliding-window rate limits, deterministic A/B routing and per-key cost tracking; contract tests bind each provider request to the real SDK signatures. |
fin-lora PyTorch PEFT · demo |
84.7% on LLM-labelled news vs 73.6% for zero-shot FinBERT (76.8% vs 71.2% on 267 human-labelled headlines). LoRA-tuned Qwen2.5-0.5B for financial sentiment; a confidence cascade cuts average latency 63% for 0.4 points of accuracy, and a follow-up study measures how easily planted labels flip it. |
| llm-safety-auditor · demo | Scores a model against the OWASP LLM Top 10; 635 tests. 50 curated attacks (350 with mutations), a 6-layer detector and PDF reports. Detector recall is measured on a seeded mock, and the README shows where it is weaker on real text. |
| readme-truth · engineering tools | Catches READMEs that no longer match the code; a hand audit of 22 repositories found 15 had drifted. Turns a README into checkable claims (test counts, versions, paths, commands, links) and checks each against the repository. Offline CLI or GitHub Action; 52 tests. |
| finance-agent-crew | Turns a stock ticker into a sourced research brief; runs offline with no API keys; 333 tests. Concurrent SEC EDGAR, market and news gathering feeds parallel fundamental, sentiment and risk agents, then a writer agent. |
| nvidia-nim-rag-techniques | Cross-encoder reranking lifted recall@1 from 0.857 to 0.964 over naive dense retrieval on a small 28-query set (too small to rank techniques). Five RAG techniques on NVIDIA NIM with LangChain and LangGraph: hybrid search, reranking, query rewriting, context compression, corrective RAG. |
| sigdet · computer vision | 0.939 ± 0.005 test mAP50-95 over 3 seeds for YOLO11n on a leakage-aware split (26 test images); the larger model scored lower and varied more. Document signature detection with a robustness study under scan corruptions. |
| exec-rl · reinforcement learning | 0.22 bps mean shortfall vs 18.8 bps for TWAP on 20,000 held-out simulated paths, 3 seeds, with an alpha signal; without the signal it is no better than TWAP. From-scratch PPO for trade execution. Simulation only. |
| guard-gap-detector | Separates all 3 vulnerable fixtures from all 5 clean ones on its synthetic corpus. Read-only static detector for a sandbox-guard gap in AST-walking restricted Python interpreters used by AI agents; artifact for the working paper When Vendoring Defeats the Patch. |
Earlier and side projects
| event-impact-signals | 94.4% (102 of 108) on held-out headlines with a sentence-embedding classifier vs 44.4% for keyword rules. The headlines are synthetic and written by one author, so this is a sanity check, not a benchmark. |
| finarena | One API in front of several fintech models: a cost-routed sentiment cascade (TF-IDF, then the LoRA LLM only when needed) and explainable credit scoring with reason codes. |
| mcp-diagram-agent | MCP server that turns a plain-text system description into an Excalidraw architecture diagram. |
| on-device-llm-optimizer | Distillation, INT4 quantization and CoreML export pipeline for running a small LLM on Apple hardware. |
| nano-finbert | A 1.88M-parameter transformer built from scratch, with a measured study of how far distillation and pretraining can take it. |
| credit-arena · trade-arena | Honest model comparisons for credit default and trading, with bootstrap confidence intervals; in trading, nothing beat buy-and-hold. |
| ledger-service | Double-entry ledger service in Go and Postgres with idempotent transfers and a concurrency test that checks balances are conserved. |
From my data, ML platform and ML engineering roles (2020 to 2026), not from the repositories above:
- Deployment failures down 85% in 60 days: built an AWS ML platform with SageMaker, an MLflow registry, gated promotion and automated rollback.
- Policy research cut from about 2 hours to under 20 minutes per question: built a RAG system over 500+ policy documents.
- Sub-200 ms p99: fine-tuned and served a 14-class BERT intent classifier to route support requests.
- 2 regulatory audits, zero findings: owned ML data security (encryption, least-privilege IAM, audit logging).
- Query runtime down 60%: led an Oracle-to-Redshift migration, rebuilding stored procedures as Airflow DAGs.
Merged upstream (20): crewAI #8061 (streaming event fix, from a bug I reported) ·
GENesis-AGI #2960 (path-containment fix) ·
prefect #23367 (ECS guide import) ·
crewAI #8033 ·
temporal sdk-python #1955 ·
ray #66918 (RLlib / Tune doc samples) ·
litestar #5126 ·
goose #1121 ·
dspy #10608 ·
jackc/pgx #2638 ·
mlflow #26613 ·
mlflow #26634 ·
langroid #1209 ·
pyproj #1637 ·
nibabel #1550 ·
pystac #1803 ·
pystac-client #931 ·
fhir.resources #211 ·
backtest-integrity-guard #33 ·
hwspec #75
Security: reported a sandbox bypass in an AI-agent code executor through a GitHub security advisory; a downstream project, femto/minion, adopted the fix and credited the report (GHSA-m5v3-6ccr-cfgx).
In review: native Groq / Fireworks / Together providers for crewAI, and fixes to broken documentation samples in Dagster, Hugging Face Datasets, Feast, LlamaIndex, sqlc and others (all open pull requests).
Agents and agent memory lethe (learned memory ranking, MCP and LangChain adapters) · finance-agent-crew (multi-agent fan-out / fan-in) · mcp-diagram-agent (MCP server)
LLM serving and routing llm-gateway (Anthropic, OpenAI, Ollama, caching, rate limits, A/B routing, cost tracking) · finarena (cost-routed model cascade)
Fine-tuning and small models fin-lora (LoRA / PEFT on Qwen2.5) · on-device-llm-optimizer (distillation, INT4, MLX) · nano-finbert (transformer from scratch)
RAG and retrieval nvidia-nim-rag-techniques (hybrid search, reranking, corrective RAG with LangGraph) · lethe (BM25 + embeddings, learned ranking)
Evaluation and LLM security llm-safety-auditor (OWASP LLM Top 10 red-teaming) · guard-gap-detector (sandbox static analysis) · credit-arena (paired bootstrap model comparison)
Computer vision and reinforcement learning sigdet (YOLO11, multi-seed evaluation) · exec-rl (PPO from scratch)
Platform and tooling fintech-devsecops-pipeline (Terraform, EKS, ArgoCD, OPA) · readme-truth (GitHub Action) · GitHub Actions CI in every repo
Languages Python · Go · SQL · TypeScript (portfolio)
Activity
Contact: nagizaazs@gmail.com · 23 public repos · Portfolio · LinkedIn · Hugging Face