
Docker Compose + Caddy TLS · GitHub Action · Stdio MCP
Zero-token static/AST pre-scans + 3-agent CrewAI Flows crew + plain-code deterministic reconciler, durable SQLite WAL queue, 8 MCP tools, and 690+ automated tests.
Hamza AliFull-Stack AI & LLM Systems Engineer — Multi-Agent Workflows, Enterprise RAG & MCP
I build the whole stack so production AI staysaccess-controlled.
Full-stack AI engineer building the whole vertical slice: FastAPI backends, React frontends, hybrid retrieval pipelines, MCP tool servers, eval harnesses, and Docker Compose deployments — ready to partner directly with product and client teams to ship AI into real workflows.
“I spend the time on the unglamorous parts: access control, tests (90 and 690+), and measuring when the model is wrong.”
690+ tests (Py 3.10–3.13) · 14-case regression suite
CrewAI Flows · Semgrep/Bandit/AST · 8 MCP tools · 15 slash cmds
90 automated tests · 0.93 RBAC · <20ms Redis cache
3-layer RBAC · ChromaDB + FTS5 RRF · DuckDB SQL sandbox
3-tier Gemini failover · 24h normalized cache
Atomic claim extraction · Google Search grounding · Electron 36
[01//flagship work]
Packaged with Docker Compose and documented with Architecture Decision Records, evaluation harnesses, and narrated walkthroughs — inspect the full code and trade-offs behind each system.

Docker Compose + Caddy TLS · GitHub Action · Stdio MCP
Zero-token static/AST pre-scans + 3-agent CrewAI Flows crew + plain-code deterministic reconciler, durable SQLite WAL queue, 8 MCP tools, and 690+ automated tests.

Docker Compose Full Stack (FastAPI + React + Redis)
3-layer RBAC (0.93 security score, 0 filter leaks), 3-way query expansion + ChromaDB/FTS5/Cohere RRF, sandboxed in-memory DuckDB SQL, <20ms Redis cache, and 90 automated tests.

Web Serverless + Electron 36 Desktop
Deconstruct articles or URLs into atomic claims, verify against live Google Search, and degrade transparently across a 3-tier Gemini failover engine.
[02//systems lab]
Step through authorization boundaries and fallback states, inspect all 8 MCP JSON-RPC tool schemas, review scoped repository benchmarks, and model LLM caching costs.
PR diff adds hallucinated package `flask-auth-helpers` + `subprocess.run(shell=True, input=user_input)`.
Step 1 / 4
Strip Unicode Bidi characters, validate HMAC/Bearer auth, and scan dependency manifests.
[03//story & experience]
The Engineer Behind the Systems
Full-stack engineer focused on AI systems that hold up outside a notebook. I build the entire vertical slice — FastAPI backends, React frontends, hybrid retrieval pipelines, evaluation harnesses, and Docker Compose deployments — so nothing gets lost in handoff between prototype and production.
Most AI demos skip the unglamorous engineering that actually decides whether a team can trust the system: per-request access control before the LLM ever runs, deterministic static scanners that constrain what agents can claim, durable webhook job queues, and automated regression tests (90 in FinSight, 690+ in AI Code Review Agent) that measure when the model is wrong.
Before focusing full-time on LLM systems, I bridged academic CS foundations and operational analytics at ATS AI Lab — turning ~1 million NOAA lightning-strike records into Tableau risk-zone recommendations — and regularly mentored fellow engineers on architecture and technical decisions. I'm looking for a full-stack or forward-deployed AI engineering role working directly with product and client teams to get AI into real workflows.
Engineering Standard
Every system ships with clear access boundaries, reproducible Docker Compose packaging, automated regression tests (90 & 690+), and honest measurement of where models fail.
Self-Directed & Open-Source Production Systems
End-to-end engineering of full-stack LLM platforms — owning FastAPI backends, React 19 TypeScript frontends, retrieval and agent orchestration, access control, automated testing (780+ tests), and Docker Compose deployment.
Advanced Telecom Services (ATS) AI Lab
Turned large-scale raw geospatial weather telemetry into stakeholder-ready operational risk intelligence alongside Microsoft Azure ML cloud workflows.
Abbottabad University of Science and Technology
Computer Science fundamentals applied directly to distributed AI systems, access-controlled data stores, and peer technical mentorship.
DeepLearning.AI
Role-scoped agent crews, provenance guardrails & tool orchestration
DeepLearning.AI
Chunking, hybrid dense+sparse retrieval, reranking & RAG evals
DeepLearning.AI
Neural network architectures, optimization, sequence models & diagnostics
Stanford Online
Supervised & unsupervised learning, bias/variance & model evaluation
[04//engineering notes]
Short architectural notes written directly from production trade-offs encountered while building FinSight, AI Code Review Agent, and Factscope AI.
Prompt instructions are not an access-control boundary. In enterprise RAG, permissions must be enforced before any embedding search, SQL query, or model call.
Many RAG prototypes retrieve top-k chunks first and tell the LLM in the system prompt to 'ignore documents the user should not see.' That pattern fails immediately under prompt injection and still leaks sensitive tokens into model context.
In FinSight, department access runs across three independent layers: JWT verification with per-request DB role re-validation, a zero-token pre-LLM department guard, and hard ChromaDB metadata filters plus an isolated in-memory DuckDB table whitelist (`enable_external_access=False`).
The automated RBAC security suite scored 0.93 with zero metadata-filter violations, while transparently surfacing a general-doc 'revenue' phrase calibration note — proving why access boundaries must be measured with regression tests, not assumed.
Key Engineering Takeaways
Sending raw diffs straight to an LLM invites hallucinated reviews. Running static scanners, AST call graphs, and a plain-code reconciler constrains agent output to verifiable facts.
LLM APIs and search grounding tools have separate rate limits. Production apps need multi-tier failover, error classification, and transparent UI state when grounding is unavailable.
[05//verified stack]
Every framework, database, and security layer listed below maps to real code in FinSight, AI Code Review Agent, Factscope AI, or ATS AI Lab — zero resume filler.
Gemini 2.5 Flash & 3.1 Flash Lite with Google Search grounding, structured outputs, fallback chains, and gemini-embedding-2-preview.
Multi-agent review crews with bounded tool loops, opt-in cross-model adversarial critique, and LCEL RAG retrieval pipelines.
Exposes SAST scans, governance rules, Tree-Sitter impacted callers, semantic context, team memory, and test generation to Claude Code & Cursor.
LLM-as-judge (RAGAS Answer Relevancy 0.894, Faithfulness 0.850; DeepEval G-Eval) paired with quota-free cosine/BM25/ROUGE-L & regression gates.
Role-partitioned ChromaDB cosine search fused with SQLite WAL/FTS5 BM25 keyword retrieval using Reciprocal Rank Fusion (k=60).
Optional second-stage cross-encoder scoring (rerank-english-v3.0) promoting the top-6 highest-signal passages with graceful RRF fallback.
In-memory DuckDB sandbox (`enable_external_access=False`, 256MB cap, 10s interrupt timer, 1,000-row limit) with automatic SQL -> RAG fallback.
Multi-tier Redis 7 + in-memory cache using SHA-256 role-isolated keys (`finsight:query:{role}:{digest}`), 300s TTL, and upload invalidation.
Typed FastAPI + Uvicorn services with Pydantic validation, SSE & NDJSON token streaming, durable SQLite job queues, and Pytest harnesses.
Cross-file symbol resolution and impacted-caller blast radius across Python, TypeScript, JavaScript, Go, Java, and Rust.
Layered static analysis, Trojan Source / slopsquatting shields, OASIS SARIF 2.1.0 export, and zero-cost `.code-review.yaml` governance.
HS256 JWT + live DB role re-fetch, SlowAPI tiered limits (/login 5/min, /upload-docs 10/min, /chat 30/min), and 20 OWASP security tests.
Live annotated code viewers, interactive SVG blast-radius graphs, progressive NDJSON chat interfaces, and Electron 36 desktop packaging.
Multi-stage Docker builds with non-root UID 10001, read-only root FS, `cap_drop: ALL`, Caddy automatic TLS, and persistent volumes.
Reusable GitHub Action, Python 3.10–3.13 CI matrix, mypy, pip-audit, Gitleaks secret scanning, Playwright E2E, and PR slash commands.
Processed ~1M NOAA lightning-strike telemetry records into Tableau risk heatmaps and operational recommendations at ATS AI Lab.
[06//collaboration]
Looking for a full-stack or forward-deployed AI engineering role where I work directly with product and client teams to get AI into real workflows — plus selective architecture & evaluation engagements.
I own the full vertical slice — FastAPI backends, React/TypeScript UIs, hybrid RAG + 3-layer RBAC, CrewAI/MCP agent workflows, 780+ automated tests, and Docker Compose deployments — working directly with engineering and client teams remotely (UTC+5).
Production knowledge workspaces enforcing department permissions before retrieval, blending dense + BM25 hybrid search, and isolating structured SQL analytics.
Deliverable: Architecture + Tested Implementation + Docker Compose Handoff
Discuss this engagementMulti-agent workflows and stdio MCP servers grounded in static/AST pre-scans, plain-code reconciliation, durable webhook queues, and fail-closed policies.
Deliverable: Agent Pipeline + MCP Server + CI/SARIF Integration
Discuss this engagementMeasure when models are wrong and harden brittle prototypes with RAGAS/DeepEval harnesses, quota-free regression suites, failover chains, and cost/latency tuning.
Deliverable: Technical Audit + Automated Eval Harness + Remediation Plan
Discuss this engagement[07//contact]
Direct email response · Remote (UTC+5)
Full-Stack AI & LLM Systems Engineer · Python, FastAPI, React & LLM Systems
Open to full-time remote Full-Stack AI and Forward-Deployed AI Engineer roles working directly with product or client teams, as well as selective consulting on RAG, multi-agent pipelines, and evaluation harnesses.