Ishtiaqe Hanif
Hamburg, Germany · open to opportunities

Software Engineer,
Applied AI research.

I’m Ishtiaqe Hanif. I spent three years at Gigalogy building retrieval and voice systems — most recently graph-RAG over knowledge graphs. Now reading for an MSc in Hamburg.

See the workGet in touch
Selected work

What the work
actually moved.

One system, measured before and after. These are the results in that codebase under its own constraints — not a promise about anyone else’s stack.

Case study · SmartAds delivery pipeline · Gigalogy, 2024–25

99.7%
Faster API response
on the worst-performing paths
10,000 → 140
API calls per operation
batching and deduplication
2h → 3min
Batch job runtime
same job, same data volume

The honest version: These came from a system with specific, findable bottlenecks — redundant calls in a loop and an unbatched job. The size of the win says as much about where the code started as about the fix. What travels is the method: measure first, find the actual hot path, and change the shape of the work rather than tuning around it.

Experience

Three years, three teams,
one compounding problem.

Each move went further up the stack — from making calls cheap, to making answers instant, to changing how the system knows things at all.

Feb 2026 – Sep 2026

Software Engineer (R&D)

Gigalogy / R&D

Rebuilt Maira’s retrieval architecture from vector RAG to graph RAG, over a purpose-built knowledge graph with a live memory layer.

Oct 2025 – Feb 2026

Software Engineer

Gigalogy / maira-api

Built realtime voice AI over customer datasets, and shipped research and deep-thinking capabilities into the Maira assistant.

Mar 2024 – Oct 2025

Software Engineer

Gigalogy / SmartAds

Performance and throughput work on the ad delivery pipeline — measured before and after, and written up as a case study.

Full timeline →
Research

Published, cited, peer-reviewed.

Machine learning applied to medical diagnostics and media forensics, with Prof. Mohammad Monirujjaman Khan at North South University.

2023Preprint

Deep Fake Analysis Using Convolutional Neural Network and Vision Transformer

Preprint · With Mohammad Monirujjaman Khan

Dual-stage deepfake detection: a fine-tuned EfficientNet extracts frame-level features, and a Vision Transformer classifies the resulting sequence for spatial and temporal manipulation artifacts.

Nov 2022Published · IEEE

Liver Cirrhosis Prediction using Machine Learning Approaches

IEEE UEMCON 2022 · With Mohammad Monirujjaman Khan

Comparative diagnostic study training SVM, Random Forest and Logistic Regression classifiers on clinical metrics to predict cirrhosis staging, with feature selection tuned for early-stage detection.