LLM, generative AI and agentic AI engineer. I build the parts of an LLM product that have to be right: the gateway in front of the models, the memory an agent reads from, the evaluation that says whether it is safe, and the ledger that records the money. Fintech first. 5 years in data and ML engineering, the last 2 in production at regulated credit unions. AWS ML Specialty certified.
Now: doctoral research (DBA, Applied AI) on AI-agent security. I traced a sandbox bypass into open-source AI-agent projects that copied the vulnerable code (details under coordinated disclosure) and earned reporter credit on GHSA-m5v3-6ccr-cfgx. Open to LLM, GenAI and agentic AI engineering roles, internship or full-time.
llm-gateway Python FastAPI Redis |
One OpenAI-compatible endpoint over Anthropic, OpenAI and Ollama with response caching, sliding-window rate limits, deterministic A/B routing and per-key cost tracking. 428 tests, 99% line coverage, CI fails under 95%. |
ledger-service Go Postgres |
Double-entry ledger: idempotent transfers, ordered row locks, append-only postings. A 40-goroutine test asserts balances are conserved; a load test held ~2,300-2,400 transfers/s (500 accounts, one laptop) with zero balance drift. |
lethe Go · demo |
Memory server for AI agents that learns from feedback which memories to return: 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, and the README says so. |
| llm-safety-auditor · demo | LLM red-teaming: 50 curated attacks (350 with mutations), a 6-layer detector, OWASP LLM Top 10 scoring, PDF reports. 635 tests. |
| fin-lora · demo | LoRA-tuned Qwen2.5-0.5B for financial sentiment: 84.7% on LLM-labelled news vs 73.6% for FinBERT (76.8% vs 71.2% on human-labelled headlines); a confidence cascade cuts average latency 63% for 0.4 points of accuracy; follow-up studies measure how easily it is fooled. |
| finance-agent-crew | Multi-agent research: concurrent SEC EDGAR, market and news gathering feeds parallel fundamental, sentiment and risk agents; a writer agent produces a sourced brief. Runs offline with no API keys. 333 tests. |
| 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. |
| guard-gap-detector | Static detector for a sandbox-guard gap in AST-walking restricted Python interpreters used by AI agents; artifact for the paper When Vendoring Defeats the Patch. |
Earlier and side projects
| mcp-diagram-agent | MCP server that turns a plain-text system description into an Excalidraw architecture diagram. |
| nvidia-nim-rag-techniques | Five RAG techniques on NVIDIA NIM: hybrid search, reranking, query rewriting, context compression, corrective RAG. |
| 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. |
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: reporter credit on GHSA-m5v3-6ccr-cfgx, a sandbox bypass in an AI-agent code executor; further reports are under coordinated disclosure.
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).
LLM serving and routing llm-gateway (Anthropic, OpenAI, Ollama, caching, rate limits, A/B routing, cost tracking) · finarena (cost-routed model cascade)
Agents and tool use finance-agent-crew (multi-agent fan-out / fan-in) · mcp-diagram-agent (MCP server) · lethe (agent memory, MCP and LangChain adapters)
RAG and retrieval nvidia-nim-rag-techniques (hybrid search, reranking, corrective RAG with LangGraph) · lethe (BM25 + embeddings, learned ranking)
Fine-tuning and small models fin-lora (LoRA / PEFT on Qwen2.5) · on-device-llm-optimizer (distillation, INT4, MLX) · nano-finbert (transformer from scratch)
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)
Backend and platform ledger-service (Go, Postgres, OpenAPI) · fintech-devsecops-pipeline (Terraform, EKS, ArgoCD, OPA) · GitHub Actions CI in every repo
Languages Python · Go · SQL · TypeScript (portfolio)
Activity
Contact: nagizaazs@gmail.com · 22 public repos · Portfolio · LinkedIn · Hugging Face