AI/ML Engineer. M.Sc. Media Engineering at TU Ilmenau, Germany.
I build LLM and deep learning systems end to end — retrieval and agent architectures, neural audio models, and the evaluation infrastructure that tells you whether either of them actually works. Three years of prior production data engineering (Razor Group Berlin, Teradata, NETSOL) is why my projects tend to be instrumented rather than demoed: latency budgets, cost per query, measured quality, and failure modes documented rather than hidden.
Currently focused on making LLM systems measurable, and on locating the real ceiling in learned audio compression.
Multi-agent RAG over public logistics operational documentation.
Fixed-orchestration agent pipeline (planner, retriever, responder) with hybrid BM25 and dense retrieval, Qdrant vector store, FastAPI backend, Streamlit UI. Instrumented end to end with LangSmith tracing and RAGAS evaluation. Measured at 0.98 faithfulness, ~9.4s median latency, $0.028 per query. 108 tests, Dockerised, CI on every push, deployed live.
Architectural decisions are documented in the repo — including why the retriever agent is deliberately not LLM-backed, and why I chose fixed orchestration over an AutoGen group chat.
Low-latency neural speech codec, rate-distortion research.
Causal transformer encoder/decoder with residual vector quantisation, targeting 8–16 kbps for real-time speech. Run as an eight-phase controlled curriculum rather than a hyperparameter sweep, including a deliberate counter-experiment (D-VAE) designed to fail. Quality plateaued at 3 bits per codebook, locating the ceiling in latent entropy rather than quantiser resolution. Known limitations — including a torch.triu/torch.tril bug that invalidated one attention-window phase — are disclosed in full.
Predicting energy decay curves from room geometry. 103M-parameter model reaching R² = 0.9995 on T20 and C50 prediction, trained on measured spatial room impulse responses.
Comparative study of PCA, LDA, and EFA for room acoustics classification. Parameter sensitivity analysis across octave bands on the Stolz et al. (2024) SRIR dataset — DT20m dominant under LDA, DRR and Grel under EFA. MATLAB.
| LLM & retrieval | Deep learning | ML engineering | Data |
|---|---|---|---|
| LangChain, AutoGen | PyTorch, torchaudio | Docker, GitHub Actions | Python, SQL, dbt |
| Qdrant, RAGAS, LangSmith | Causal CNNs, transformers | pytest, CI/CD | Spark, Airflow |
| FastAPI, Streamlit | RVQ, autoregressive models | AWS, Azure | Snowflake, Teradata |
| Hybrid BM25 + dense retrieval | librosa, PESQ, STOI | Model evaluation & tracing | Power BI, MATLAB |
TU Ilmenau — M.Sc. Media Engineering, Faculty of Electrical Engineering and Information Technology. In progress.
Razor Group GmbH, Berlin — Data Engineer. E-commerce operations and marketplace analytics.
Teradata, Islamabad — Enterprise data warehousing.
NETSOL Technologies, Islamabad — Telecom BI.
B.Sc. Software Engineering.
Microsoft DP-900 · Teradata Vantage
English (C2) · German (B2) · Urdu, Punjabi (native)