Umair Talib

Umair Talib

Applied AI engineer for LLM applications, speech AI and model evaluation.

After over 1.5 years as a software engineer, I'm developing AI products part-time: an AI platform for public-funding workflows at ATTEC, as its only developer, and an AI-assisted planning tool for municipal citizen participation at the University of Siegen. I make the design decisions and write the specs, AI coding agents implement, and I review and test every change. I'm an M.Sc. Computer Science student and co-author of a conference paper on removing interviewer speech from clinical interview recordings, published by IEEE (CoDIT 2026).

CV (PDF)Paper (DOI)

Results

Every number here comes with what it was measured on.

Tuning sets are called tuning sets, test sizes are given, and confidence intervals stay attached. Where the code is private, the source says so.

ResultWhat it measures, and on whatDocumented in
14.18% → 1.09%interviewer contamination in participant transcriptsInterviewer prompts left in the participant transcripts, before and after the pipeline, as reported in the published paper.CoDIT 2026 Published paper (DOI)
0invented or wrong answers from the grounded extractionOn three evaluation sets used for tuning, 113 cases. No unseen data. Tag recall 98.5%.AI platform for public-funding workflows Evaluation sets in the private repository
0.614ROC-AUC on the held-out test56 participants, 95% CI 0.454 to 0.771. Recall 88.2% at precision 41.7%. Screening for elevated depressive symptoms (PHQ-8 of 10 or more).Research project Research project report, University of Siegen
34.1 h in ≈ 2.5 hinterview audio transcribedOn 10 GPU nodes of the university HPC cluster with SLURMResearch project Research project report, University of Siegen
≈ 3,700automated testspytest, Vitest and Playwright, including property, golden and mutation testsAI platform for public-funding workflows Private repository
≈ 9,000automated testspytest, Vitest and Playwright, including a permission matrix that fails when a route has no rights entryFUSION toolbox Private repository
0.976 – 0.989best validation AUROCs of 30 fine-tuned networksValidation set of the course projectCourse project Course project, University of Siegen

Work

Two AI products in development, built with AI coding agents and held to tests.

06/2026 – present ATTEC UG (haftungsbeschränkt) Siegen, hybrid

AI platform for public-funding workflows

Software Engineer, AI Product (part-time)

An AI platform that aims to automate public-funding workflows for German SMEs and the funding experts who advise them, from finding programmes to the application and managing the funded project. I am its only developer. The first stage already works: it shows which programmes from a curated, growing catalogue fit a company, with every checked condition, its reason and the official source.

Code decides, the LLM only reads language
  1. inputCompany texts and websites
  2. LLMThe LLM extracts facts and must cite the words it used
  3. codeCode verifies each cited word in the input
  4. personThe user confirms the facts
  5. codeA rules engine checks each programme's conditions
  6. outputEach result lists every checked condition, its reason and the official source

solid: codedashed: LLMdotted: person

  • Built a deterministic Python rules engine. Code decides and the LLM only reads language, so the same input always gives the same result, and each result replays from its data version, engine version and date.
  • Designed grounded LLM extraction. The model must cite the words it used, code verifies them in the input, and the user confirms.
  • Model each programme's conditions as yes/no facts and checks in a pydantic-validated YAML model. Every fact needs a verbatim quote from a saved official text, and data tests check the quotes, numbers and dates.
  • About 3,700 automated tests (pytest, Vitest, Playwright) with property, golden and mutation tests, and GitHub Actions CI with secret scanning and dependency audits.
  • Each feature runs through spec, build with AI coding agents, three independent AI reviews and fixes, my own review and tests, and written decision records.

In development. The code is private, so it is described here in words only.

07/2026 – present University of Siegen, research project FUSION Siegen, hybrid

FUSION toolbox

Student Research Assistant (WHB), Full-Stack Development (part-time)

An AI-assisted planning tool for municipal citizen participation. It derives a project profile from a description and documents, recommends participation formats and turns the choice into a roadmap with budget, tasks and calendar.

Every AI task has a fallback without the LLM
  1. inputProject description and documents
  2. codeDeterministic filters and a fit score select candidate formats
  3. LLMAn LLM rates the candidates against format profiles, with quoted reasons
  4. codeFixed rules check the LLM's answer
  5. outputA roadmap with budget, tasks and calendar

solid: codedashed: LLM

  • A collection of test cases checks the recommendations, and every AI task has a fallback that works without the LLM.
  • Modular monolith with 15 domain modules. FastAPI (Python 3.13), SQLAlchemy 2, Alembic, PostgreSQL 16 as the only data store (jobs and events in tables) and a background worker. Front end in Next.js 16, React 19 and TypeScript with a client generated from OpenAPI.
  • Built with AI coding agents in a spec, plan, build and review cycle per sub-project. About 9,000 automated tests (pytest, Vitest, Playwright), including a permission matrix that fails when a route has no rights entry, and GitHub Actions CI.
  • Earlier, built the first version (07–08/2026) with an AI coding agent (Antigravity): accounts, roles and a Kanban board on FastAPI, PostgreSQL and Next.js. Reviewed and approved the architecture of the current version (09/2026).

In development. The code is private, so it is described here in words only.

Before the Master's

02/2022 – 09/2023

Software Engineer, ZweiDevs

Built React and Next.js front ends and Node.js and Express back ends for several projects, including an EdTech platform and a Web3 game (React front end, game mechanics in Unity).

Implemented Stripe payments, JWT-based login and role-based access for the EdTech platform, deployed services on AWS (EC2, S3) with PostgreSQL and MongoDB, and tested with Jest and Cypress.

Method

I make the design decisions and write the specs. AI coding agents implement, and I review and test every change.

At ATTEC every feature goes through this cycle. The FUSION toolbox uses a similar spec, plan, build and review cycle for each sub-project.

  1. 01SpecI make the design decisions and write a specification for each feature.
  2. 02BuildAI coding agents implement the spec, with tests.
  3. 03AI reviewsThree independent AI reviews read the change, and what they find gets fixed.
  4. 04My reviewI review the change and test it myself.
  5. 05RecordDecisions are written down as decision records.
Tests
Automated tests with pytest, Vitest and Playwright, including property, golden and mutation tests.
Eval sets
Evaluation sets gate prompt changes at ATTEC.
Rights
In the FUSION toolbox, a permission matrix fails when a route has no rights entry.
CI
GitHub Actions, with secret scanning and dependency audits at ATTEC.

Research

Does the model use the right evidence?

Speech and multimodal machine learning for mental-health screening, with a focus on whether a model uses the right evidence: leakage-resistant data pipelines and the evaluation of grounded LLM outputs.

Conference paper · CoDIT 2026, published by IEEE

Leakage-Resilient Diarization and Feature Extraction Framework for Robust Multimodal Mental Health Corpora

U. Oleh, U. Talib, M. A. Alam, and R. Obermaisser. In Proc. 12th Int. Conf. Control, Decision and Information Technologies (CoDIT), Bari, Italy, Jul. 2026, pp. 2508–2512.

In clinical interview recordings, the interviewer's speech can end up in the participant's transcripts, and a model may then learn from the interviewer's prompts instead of from the participant. The paper describes a speaker-verified diarization and transcription pipeline with a text filter for interviewer prompts, which I built in my Master's research project, shown below.

Interviewer contamination in participant transcriptsInterviewer prompts left in the participant transcripts, before and after the pipeline, as reported in the published paper.
Read it on IEEE Xplore doi:10.1109/CoDIT70676.2026.11630977
Show BibTeX
@inproceedings{oleh2026leakage,
  author    = {Oleh, Ugonna and Talib, Umair and Alam, Muhammad Ali and Obermaisser, Roman},
  title     = {Leakage-Resilient Diarization and Feature Extraction Framework for Robust Multimodal Mental Health Corpora},
  booktitle = {Proc. 12th Int. Conf. Control, Decision and Information Technologies (CoDIT)},
  address   = {Bari, Italy},
  year      = {2026},
  month     = jul,
  pages     = {2508--2512},
  publisher = {IEEE},
  doi       = {10.1109/CoDIT70676.2026.11630977}
}

04/2025 – 03/2026 · University of Siegen, Chair for Embedded Systems

An Audio-Dominant Framework for Robust Depression Detection and Clinical Explainability

Module Project Work, grade 1.0

Screening for elevated depressive symptoms (PHQ-8 of 10 or more) from clinical interview recordings, with a gated fusion model on speech and text embeddings.

  • Built a speaker-verified diarization and transcription pipeline (pyannote.audio, Faster-Whisper) with a text filter for interviewer prompts. In the published paper, interviewer contamination in participant transcripts fell from 14.18% to 1.09%.
  • Transcribed 34.1 hours of clinical interview audio in about 2.5 hours on 10 GPU nodes of the university HPC cluster with SLURM.
  • Trained a gated fusion model on WavLM and Mental-RoBERTa embeddings with a session-level Transformer on top.
  • Built an explanation pipeline that turns attention-based evidence into text explanations with an LLM, shown in a research dashboard in Streamlit.

0.614

ROC-AUC on the held-out test

56 participants, 95% CI 0.454 to 0.771. Recall 88.2% at precision 41.7%. Screening for elevated depressive symptoms (PHQ-8 of 10 or more).

34.1 h in ≈ 2.5 h

interview audio transcribed

On 10 GPU nodes of the university HPC cluster with SLURM

08/2026 · Course project, University of Siegen, with one co-author

NAS-Bench-201 transfer to cosmic-ray detection

Transferred NAS-Bench-201 architectures to cosmic-ray detection in telescope images. Thirty fine-tuned networks reached best validation AUROCs between 0.976 and 0.989.

Timeline

Lahore to Siegen.

Education

  • B.Sc. (Hons) Computer Science10/2017 – 03/2022 · University of Central Punjab, Lahore

  • Artificial Intelligence course10/2021 – 12/2021 · Samsung Innovation Campus, Lahore

  • M.Sc. Computer Science10/2023 – present · University of Siegen, current average 2.1

Work

  • Software Engineer02/2022 – 09/2023 · ZweiDevs

  • Software Engineer, AI Product06/2026 – present · ATTEC UG, part-time

  • Student Research Assistant (WHB), Full-Stack Development07/2026 – present · University of Siegen, part-time

Research

  • Research project, depression screening from speech04/2025 – 03/2026 · Chair for Embedded Systems, grade 1.0

  • Paper at CoDIT 2026, published by IEEE07/2026 · Bari, Italy

  • Course project, NAS-Bench-201 transfer08/2026 · with one co-author

Skills

Skills and languages

AI and ML
PyTorch, Hugging Face Transformers, WavLM, Whisper, pyannote.audio, LLM APIs with structured output, evaluation sets
Backend
Python, FastAPI, pydantic, SQLAlchemy, Alembic, PostgreSQL, REST, JWT authentication
Front end
TypeScript, React, Next.js, Tailwind CSS
Engineering practice
pytest, Vitest, Playwright, GitHub Actions, Docker Compose, Ruff, mypy, Git
Compute
SLURM on an HPC cluster with A100 GPUs
AI coding agents
Claude Code and Antigravity, controlled by tests, evaluation sets and decision records
Languages
English, C1, fluent (M.Sc. programme taught in English)

Contact

Write to me.

I'm happy to talk about AI engineering: LLM applications, speech AI and model evaluation.

Send an e-mail

Profiles

CV as PDF

Based in

Siegen, Germany