About
I am a fully independent researcher. I work on researching open problems in deep learning and optimization. I like problems where I can write the model and the math from scratch rather than call an API.
Publications
A Spectral Bound on Effective Sharpness for Fisher-Preconditioned Gradient Descent
Neural networks trained with gradient descent exhibit the Edge of Stability (EoS), where Hessian sharpness rises toward 2/η and the loss oscillates non-monotonically. This paper asks whether that instability persists under Fisher-preconditioned optimization. We analyze effective sharpness, Seff = λmax(F⁻¹H), and prove the general bound Seff ≤ 1 + (ϵ + δ)/µmin(F), where ϵ = ‖H − G‖₂ is residual curvature, δ = ‖G − F‖₂ is the Gauss-Newton/Fisher gap, and µmin(F) is the minimum Fisher eigenvalue. The idealized G = F case yields the simpler bound 1 + ϵ/µmin(F), explaining natural gradient descent's spectral-flattening mechanism while remaining a special case. A stochastic extension bounds mini-batch effective sharpness under sampling noise and holds in all 200 tested draws. An alignment-aware Rayleigh analysis explains why empirical-Fisher worst-case bounds can be loose: residual spectral mass aligns with better-conditioned Fisher directions. Exact deep linear network experiments from 55 to 3,240 parameters verify the general bound across all tested scales; 110-parameter exact-matrix traces show the idealized bound fails when the empirical-Fisher gap is substantial. Matrix-free estimates extend δ, ϵ, and Seff measurement to MNIST and CIFAR-10 ResNet-18, revealing a stable relative empirical-Fisher gap of about 0.70–0.75 at 11.2M-parameter scale. On CIFAR-10, K-FAC reaches 90.5% test accuracy, while tuned first-order and adaptive baselines show that approximation quality and hyperparameter selection determine practical convergence.
@article{ahad2026a,
title = {A Spectral Bound on Effective Sharpness for Fisher-Preconditioned Gradient Descent},
author = {Abdul Ahad and Napa Lakshmi},
journal = {Transactions on Machine Learning Research},
issn = {2835-8856},
year = {2026},
url = {https://openreview.net/forum?id=EabuvggEbb}
}
Projects
Kivixa Productivity Workspace
Flutter, Dart, Rust, llama.cpp
Architected an offline, cross-platform workspace distributed via F-Droid and winget, powered by a Rust-native AI engine and local-first design.
- Architected an offline, cross-platform workspace (distributed via F-Droid and winget) driven by a Rust-native AI engine. Engineered multi-model LLM inference with Vulkan/Metal GPU acceleration and integrated a sandboxed Model Context Protocol (MCP) for secure, autonomous file operations.
- Built a Rust-backed audio intelligence pipeline featuring Whisper STT for real-time word-level transcription, Kokoro neural TTS, and voice activity detection; integrated alongside a local vector database for semantic search and a high-performance math engine for multivariable calculus and hypothesis testing.
- Implemented Git-like version control utilizing SHA-256 content-addressable blob storage with automatic snapshots. Designed a Lua 5.3 scriptable plugin architecture with a programmatic API, an interactive force-directed knowledge graph, and parallel state-driven orchestration for chained routines.
Cryptex: Unsupervised Cryptanalysis Engine
Numba, MCMC, HMM
Engineered a multi-language cryptanalysis suite with adaptive probabilistic solvers and rigorous empirical validation for classical cipher families.
- Engineered a multi-language cryptanalysis suite supporting 7 classical cipher families, implementing adaptive MCMC (with parallel tempering), Baum-Welch HMM/EM, and KPA-constrained solvers optimized over Kneser-Ney smoothed 5-gram language models.
- Formulated a statistical cipher-type detector (leveraging Index of Coincidence, Kasiski examination, and entropy) alongside a phase-transition analyzer that empirically maps ciphertext length against decryption success rates to validate theoretical unicity distances.
- Executed a rigorous 130-run empirical benchmarking framework, achieving a 100% success rate and 0.0 Symbol Error Rate (SER) on substitution and transposition ciphers by accelerating the evaluation hot loop with Numba JIT-compiled tensor operations.
Universal Adversarial Cloak
PyTorch, Hugging Face, FaceNet, CLIP
Developed an adversarial optimization framework to cloak identity signals against biometric and semantic recognition models while preserving visual fidelity.
- Engineered an adversarial optimization engine using Projected Gradient Descent (PGD) to mathematically cloak biometric and semantic identities against FaceNet and CLIP architectures.
- Enforced strict L-infinity norm constraints within the joint loss function to guarantee human-imperceptible perturbations, achieving a structural similarity index (SSIM) of > 0.98 across all cloaked outputs.
- Validated attack efficacy via an automated offline benchmarking pipeline, proving high black-box transferability by successfully blinding industry-standard oracle models (ArcFace, CLIP ViT-L/14) and optimizing Mean Residual Similarity (MRS) through comprehensive ablation studies.
GPU-Accelerated Geometric Image Reconstruction
PyTorch, NumPy, CUDA, OpenCV
Built a high-throughput geometric reconstruction system that approximates photographic images with optimized primitive-based composition under strict compute budgets.
- Architected a GPU-accelerated sequential hill-climbing engine to iteratively reconstruct high-fidelity photographic images using geometric primitives, employing a coarse-to-fine multi-stage schedule to capture both foundational structures and micro-details.
- Formulated an analytic closed-form color solver and integrated Sobel-based gradient maps for structure-aware shape routing, completely eliminating stochastic color search loops and maximizing algorithmic convergence speed.
- Executed rigorous empirical ablation studies across 11 architectural variants, achieving a peak Structural Similarity (SSIM) of 0.6059 and PSNR of 17.96 dB within a strict 1-minute compute budget, significantly outperforming standard evolutionary baselines.
Signature Verification System with Explainable AI
Siamese Neural Network for offline authentication
Determines signature authenticity by comparing a query signature against a known reference - entirely offline via metric learning.
- Siamese Architecture: Dual-branch CNN sharing weights to map signatures into a common embedding space.
- Metric Learning: Contrastive / Triplet Loss to minimise distance between genuine pairs and maximise it for forgeries.
- Preprocessing Pipeline: Normalisation, binarisation, and noise reduction for robust feature extraction.
Hospital Operations System
Flask, Vue.js, Celery, Redis, Docker, Prometheus
Architected a dual-module hospital and blood bank platform with secure API design, asynchronous workflows, and production observability.
- Architected a dual-module hospital and blood bank platform deployed via Docker Compose with a non-root Gunicorn/tini runtime. Secured the API with dual Session/JWT authentication, Argon2 hashing, and strict validation (Pydantic v2, Flask-Limiter, Talisman CSP).
- Engineered an SQL-heavy allocation engine with trigger-backed forensic auditing and compatibility matching. Implemented a Redis/Celery async pipeline for scheduled workflows and PDF/Pandas CSV exports, supported by SQLite WAL tuning to handle concurrent booking contention.
- Integrated production-grade observability featuring structured JSON logging (structlog) with cross-request X-Request-ID tracing, RFC 7807 error contracts, dependency health probes, and custom Prometheus business metrics visualized in Grafana.
Skills
| Languages | Python Rust Dart SQL |
| Deep Learning & Generative AI | PyTorch Hugging Face Transformers PEFT/LoRA llama.cpp Unsloth |
| Backend | FastAPI Node.js Celery Redis Docker AWS |
| Databases & Observability | PostgreSQL SQLite MongoDB Prometheus Grafana |
| Cross-Platform | Flutter Tauri |