Skip to content
View bakhtawar-i's full-sized avatar

Achievements

Achievement: Pull Sharkx2Achievement: YOLO

Achievements

Achievement: Pull Sharkx2Achievement: YOLO

Block or report bakhtawar-i

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.

Content in all repositories owned by your account will be closed.
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
bakhtawar-i/README.md

Hi, I'm Bakhtawar 👋

ML Engineer, applying diagnostic methods across medical AI and robotics | Open to Relocation

🎯 What I Do

I look at cases where a model appears to work and doesn't — recent projects span seizure detection from EEG and grounding failures in a robot manipulation policy. My background is in shipping production ML systems; my current focus is figuring out why they break.

  • Medical AI Diagnostics (clinical signal analysis, seizure detection)
  • VLA & Agentic System Diagnostics
  • Reinforcement Learning
  • Applied ML in Production

🔬 Research

eeg-seizure-fusion-probe (in progress) EEG seizure detection on CHB-MIT, benchmarking a feature-based GBM baseline against fine-tuned EEG foundation models (LaBraM, BENDR, NeuroGPT). Baseline GBM with expanded spectral/wavelet/cross-channel features reaches AUC-PR 0.174 (sensitivity 0.34 @ 34 FA/hr) on a held-out patient split. Currently fine-tuning LaBraM on CHB-MIT to test whether documented epilepsy-detection underperformance in foundation models replicates under a from-scratch pipeline; BENDR and NeuroGPT comparisons planned next.

vla-grounding-probe Mechanistic layer-wise diagnostic of a vision-language-action model on manipulation tasks. Found language embedding similarity running inverse to grounded conditioning predictions in a majority of tested tasks — validated against CLIP-family affirmation bias literature.

Action Space Design and Policy Optimisation for RL-Based UAV Collision Avoidance (MSc thesis, under review — Elsevier) Controlled factorial study (PPO vs. SAC × 3 action spaces × 3 reward densities, 54 runs) diagnosing SAC's catastrophic failure on hybrid action spaces, traced to its entropy auto-tuning mechanism.

Multimodal Agentic RL System FiLM vs. concat fusion comparison in MiniGrid — naive concatenation caused language to be ignored (55% instruction-following), FiLM improved to 67–90%, but 0% on novel unseen instructions. Confidence scores found uninformative.

🚀 Applied Work

Lexis — Production-grade AI document intelligence platform combining a full RAG pipeline, conversational research agent, and policy document assistant for government/enterprise use.

Sakeenah — AI-assisted journaling platform (founder), live at sakeenah-journal.vercel.app, ~20 active users.

QuantVol — Time-series forecasting system for equity volatility, benchmarking econometric models against ML/deep learning approaches.

🛠️ Tech Stack

  • ML & Modeling: PyTorch, Scikit-learn, XGBoost, TensorFlow, Stable-Baselines3
  • LLM & Agentic Systems: LangChain, LangGraph, LangSmith, RAG, Prompt Engineering
  • Vision & Robotics: SmolVLA, LIBERO, YOLO, OpenCV, ROS2, Gazebo
  • MLOps: FastAPI, Docker, AWS, GCP, MLflow, CI/CD
  • Data: Pandas, NumPy, SPSS, Power BI

📝 Publications

  • Iftikhar, B., Ahmad, R., Ghafoor, S., & Qazi, S. Action Space Design and Policy Optimisation for Reinforcement Learning-Based UAV Collision Avoidance in Post-Earthquake Environments. Under review, Elsevier.
  • Iftikhar, B. et al. (2020). Cost-effective, Reliable, and Precise Surface Mount Device (SMD) on PCBs. IOP Conference Series: Materials Science and Engineering.

🎯 Currently

Focused on exploring opportunities in health-tech and medical AI. Also open to research assistant roles and AI strategy consulting.

📫 Connect

Interested in why AI systems fail quietly, not just whether they work.

Pinned Loading

  1. EEG-Seizure-Fusion-Probe EEG-Seizure-Fusion-Probe Public

    EEG seizure detection on CHB-MIT, benchmarking a feature-based GBM baseline against fine-tuned EEG foundation models (LaBraM, BENDR, NeuroGPT). Baseline GBM with expanded spectral/wavelet/cross-cha…

    Python

  2. Postoperative-Sepsis-EWS Postoperative-Sepsis-EWS Public

    Postoperative sepsis early-warning system built on MIMIC-IV — Sepsis-3 labeling from first principles, qSOFA/NEWS2 baselines, and a nurse's-station alert dashboard. Research prototype, not a clinic…

    Python

  3. Chest-CT-ZeroShot-Benchmark Chest-CT-ZeroShot-Benchmark Public

    Zero-shot chest CT abnormality classification: comparing a 3D-native vision-language model (CT-CLIP) against a 2D slice-pooled baseline (BiomedCLIP) on CT-RATE, with a calibration analysis across 1…

    Python

  4. Astro-IB-Probe Astro-IB-Probe Public

    Information Bottleneck analysis of a self-supervised astronomical foundation model, validated with three MI estimators and a positive control.

    Python

  5. Multimodal_Agentic_RL_System Multimodal_Agentic_RL_System Public

    This project aims to build a multimodal, instruction-conditioned reinforcement learning (RL) agent, extend it with agentic reasoning using LLMs, evaluate it for generalization and robustness, and d…

    Python

  6. Lexis Lexis Public

    A production-grade AI document intelligence platform built for enterprise and government use cases. Lexis combines a full RAG pipeline, a conversational research agent, and a policy document assist…

    Python