Skip to content
View ejazfahil's full-sized avatar
🎯
Focusing
🎯
Focusing

Block or report ejazfahil

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
ejazfahil/README.md

Fahil Ejaz

AI/ML Engineer · Data Scientist — production RAG agents, LLM fine-tuning & serving, MLOps, and rigorous statistical evaluation.

LinkedIn GitHub Focus


About

I build production-grade AI systems that bridge research and deployment — stateful RAG agents, QLoRA fine-tuning pipelines, real-time drift monitoring, and statistically sound model evaluation. My work favors honest benchmarks, maintainable code, and observable systems.

  • 🔭 Currently: production RAG & agents, QLoRA fine-tuning, MLOps & drift detection
  • 🧪 Specialties: LLM/VLM evaluation, statistical inference, data engineering
  • 🏭 Domains: Industrial AI, NLP, anomaly detection, reinforcement learning
  • 🎯 Open to: AI Engineer · ML Engineer · Data Scientist · Research roles

On this portfolio: every project README reports only what its code and committed artifacts can actually prove. Real outputs are cited as numbers; scaffolds and blueprints are labeled as such. No inflated metrics.


⭐ Featured Projects

Project What it is Stack
Production RAG + Human-in-the-Loop Agent Stateful agent for industrial document intelligence — LangGraph state machine with a human approval gate (retrieve → draft → grounding-check → approve → finalize/abstain), pgvector retrieval, durable checkpointing, Prometheus metrics, and a Postgres audit trail. MIT-licensed, strictly typed. LangGraph · pgvector · FastAPI · Prometheus
QLoRA → Quantize → Serve End-to-end 7B fine-tuning pipeline: QLoRA (4-bit NF4) → eval → merge → AWQ → vLLM serving. Config-driven, reproducible, transparent eval harness (GPU benchmarks marked pending — no placeholder numbers). PEFT · TRL · bitsandbytes · vLLM
Vectorless RAG Lab Research harness comparing 7 embedding-free retrieval pipelines — tree-navigation, BM25, agentic search, a hybrid RRF router, quote-extraction, and a novel three-stage hybrid. Local-first LLM client, telemetry, and a RAGAS / LLM-as-judge eval scaffold. BM25 · RAGAS · Ollama
AI for Business — 3 Case Studies Churn, energy forecasting & segmentation with reproduced results: churn ANN ≈ 84% test accuracy, energy demand Random Forest R² ≈ 0.68, segmentation silhouette ≈ 0.41 (k=2). TensorFlow · scikit-learn · pandas
Bayesian LLM/VLM Evaluation Goes beyond point estimates: partial-pooling logistic model (correct ~ system + (1|field) + (1|doc_class)) via Bambi/PyMC, posterior contrasts with 94% HDIs, ArviZ diagnostics & LOO. PyMC · Bambi · ArviZ
Doc Extraction Benchmark (VLM vs OCR) Honest, pre-registered field-level extraction benchmark on CORD — Pixtral vs Tesseract/EasyOCR, normalized exact-match scoring, document-clustered bootstrap CIs + exact McNemar tests. Pixtral · Tesseract · pandera · SciPy

🏗️ Data & ML Engineering Suite

Focused, production-shaped building blocks — each a compact but working implementation.

Repo What it does Stack
data-quality-framework Schema / null / range / freshness / SLA validators with HTML reports & alerts pandas · pytest
airflow-etl-pipeline Production ETL DAGs: CSV→Postgres, API ingest, dbt trigger, Slack alerts Airflow
dbt-analytics-models Staging → intermediate → mart models with schema tests & freshness SLAs dbt · SQL
kafka-event-streaming Idempotent producers, manual-commit consumers, DLQ, replay, lag monitor Kafka
spark-streaming-kafka PySpark Structured Streaming: watermarked windowed aggregations + JDBC sink PySpark
postgres-data-modeling Star schema, range partitioning, BRIN/partial indexing, Alembic, pgTAP PostgreSQL

🔬 Research & Applied ML

Repo Focus
mlops-drift-detector Real-time data/concept drift — PSI + Page-Hinkley, streaming monitor, Prometheus exporter
moon_lander_rl Deep RL — DQN agent (PyTorch) on LunarLander-v3, with checkpoints, rollout videos & training curve
ALS_Disease_Severity Clinical ML — ALSFRS-R severity stratification from biomarker/functional features
OCR_Vision_Model_for_Industries Modular industrial OCR framework — ensemble OCR + LLM verification, CER/WER metrics
genai-eval-framework LLM eval harness — parallel MMLU evaluator, Anthropic/Ollama adapters, on-disk caching

🛠️ Tech Stack

Languages & Core

Python SQL PyTorch TensorFlow scikit-learn

LLM & GenAI

LangChain Hugging Face Ollama PEFT / QLoRA vLLM PyMC

Data & MLOps

PostgreSQL Kafka Spark Airflow dbt Docker Prometheus Grafana


Building production AI systems that work in the real world · Open to AI/ML Engineering, Data Science & Research opportunities · linkedin.com/in/fahil-ejaz

Pinned Loading

  1. Adaptive_Probabilistic_Testing_of_DNNs_LLMs Adaptive_Probabilistic_Testing_of_DNNs_LLMs Public

    Python 1

  2. OCR_Vision_Model_for_Industries OCR_Vision_Model_for_Industries Public

    Python 1

  3. ag_vectorless_RAG ag_vectorless_RAG Public

    vectorless RAG

    Python

  4. bayesian-llm-eval bayesian-llm-eval Public

    Hierarchical Bayesian evaluation of LLM/VLM outputs — posterior credible intervals & partial pooling over any eval's item-level scores (Bambi/PyMC).

    Python

  5. moon_lander_rl moon_lander_rl Public

    Python

  6. QLoRA-Fine-Tune-Quantization-served-with-vLLM QLoRA-Fine-Tune-Quantization-served-with-vLLM Public

    xtend my existing LLaMA 2 work into a clean, benchmarked fine-tuning project

    Python