Applied AI Engineer — I build LLM systems, ship them to production, and work to make their outputs more reliable.
💼 Open to remote Applied AI / LLM / GenAI Engineer roles
- 🏆 IBM Granite Generative AI Hackathon — Winner (Team PolyEns)
- 🚀 14 AI systems reported across 17 international hackathons, with 8 live Hugging Face demos
- 🎓 2× Stanford Code in Place Section Leader (2025 & 2026) — reported as 1 of 100 selected from 900+ global applicants
- 🏆 BICON Startup Incubation — Selected for final-year project · Proof
- 📍 Pakistan · Open to remote Applied AI / GenAI / LLM engineering opportunities
I aim to make AI systems useful, testable, and more reliable—not just impressive in a demo.
| Domain | Example work |
|---|---|
| RAG systems | Retrieval-augmented Q&A, groundedness critique, citation validation, compliance retrieval |
| AI agents | Multi-agent lead qualification, debate and reconciliation patterns, workflow agents |
| Document intelligence | PDF-to-structured extraction, resume analysis, contract fit-gap analysis |
| LLM applications | Public Hugging Face demos using Groq, LLaMA, IBM Granite, and inference APIs |
1. ComplianceRAG — RAG with a Groundedness Critic
Many RAG demos show a fluent answer. This project explores how to check whether the answer is supported by retrieved evidence.
Problem: Retrieval-augmented generation can produce plausible answers that are not adequately supported by source documents.
Approach: A pipeline combining retrieval (FAISS + all-MiniLM-L6-v2), answer generation with the Hugging Face Inference API, and a critic for sentence-level groundedness and citation validation. Low-confidence results can be flagged for human review.
Stack: Gradio · sentence-transformers · FAISS · Hugging Face Inference API · PyPDF
🔗 Live Demo · 💻 Source Code
2. AdvancedLeadsGeneration-AI — 🏆 IBM Granite Hackathon Winner
A multi-agent approach to lead qualification that weighs opportunity and risk.
Problem: Static lead-scoring rules can miss context in historical deals and customer negotiations.
Approach: A multi-agent debate workflow: one agent advocates for a lead, another surfaces risks, and a reconciliation layer produces a combined assessment using historical deal context. Built with IBM Watson AI and IBM Granite models.
Stack: Next.js · FastAPI · IBM Watson AI · IBM Granite
🏆 Hackathon Page · 💻 Source Code
3. HireMind-AI — LLM-Powered Career Workflow
A multi-feature career tool rather than a single-endpoint demo.
Problem: Candidates often tailor resumes without a clear way to compare requirements across roles or identify gaps.
Approach: An integrated workflow covering multi-job-description analysis, ATS-style scoring, skill-gap detection, resume optimization, cover-letter generation, interview preparation, application tracking, and career chat.
Stack: Streamlit · Groq API (LLaMA 3.1) · JSON persistence · Hugging Face Spaces
🔗 Live Demo · 💻 Source Code
🔬 Other Work: SafeLite — Research-Oriented AI System (click to expand)
SafeLite is a modular system exploring planning, safety reasoning, execution, simulation, and evaluation as separately testable components.
- Areas of interest: separation of planning and execution, explicit safety checks, controlled simulation, and testing component behavior and failure cases.
- Stack:
Python·FastAPI·LangChain· custom evaluation harness
- IBM Granite Generative AI Hackathon — Winner (Team PolyEns; AdvancedLeadsGeneration-AI)
- Harvard CS50x Puzzle Day 2025 — 1st Place (9/9)
- Stanford Code in Place — Section Leader (2025 & 2026)
- AlgoVerse Research Program (EOE.ai × Omdena) — accepted with a 45% merit scholarship; RAG/NLP research
- BICON Startup Incubation — Selected for final-year project · Proof
- Hackathons: 17 events · 14 submissions — LabLab profile
| Category | Technologies |
|---|---|
| Languages | Python · SQL · TypeScript |
| Backend | FastAPI · REST APIs |
| Data | PostgreSQL · Pandas · SQLAlchemy |
| LLMs & APIs | Groq (LLaMA 3.1/3.3) · IBM Granite · IBM Watson · Hugging Face Inference · OpenAI |
| RAG & retrieval | FAISS · sentence-transformers (all-MiniLM-L6-v2) · LangChain · PyPDF |
| Agent patterns | Multi-agent debate · Critic/verifier patterns · CrewAI |
| Frontend & demos | Streamlit · Gradio · Next.js |
| Deployment | Hugging Face Spaces · Streamlit Cloud |
- Ship, then evaluate. A working demo is a starting point; testing reveals where it fails.
- Retrieval quality matters. Some apparent prompting problems are actually retrieval or context-quality problems.
- Make claims testable. A feature existing in code does not mean it has been verified.
- Expose uncertainty. Low-confidence outputs should be identified rather than presented as unquestionable facts.
- Document limitations. Clear constraints make projects more useful to reviewers and contributors.
Open to Applied AI, LLM, Python backend, and AI application engineering opportunities, as well as open-source collaboration and discussions about RAG, agents, and evaluation.
- 📧 Email: faraz.outreach8@gmail.com
- 💼 LinkedIn: farazmubeenhaider
- 🤗 Hugging Face: Faraz618
- 🧪 Hackathons: LabLab
- ✍️ Writing: Medium
Build carefully. Verify honestly. Document what matters.