shahad.
About

I build AI systems that have to be right.

I'm Shahad Mahmud, an AI engineer from Bangladesh. I lead the AI team at Delineate (YC W25), where we build LLM and agentic systems for pharmaceutical research. There, an answer that sounds right but isn't is worse than no answer, so a lot of my time goes into evals and the engineering that makes models dependable in production.

I came to this through Bangla NLP. My first years were speech recognition, language-model pre-training and the lemmatization and named-entity work that became papers at EMNLP 2023 and NAACL 2025: building for a language most tools don't serve well.

I still ship code, and I write here about what I learn shipping these systems.

Shahad Mahmud, arms crossed, leaning against a red brick wall

01Career

  1. Oct 2024 – present

    Delineate (YC W25)

    Sr. AI Engineer & Team LeadPreviously Senior AI Engineer, part-time from Feb 2024
    • Lead a 10+ person AI team inside a 20+ person tech org: set technical direction, mentor engineers and still ship production code.
    • Rebuilt the document-parsing pipeline on vLLM and Kubernetes: 45% lower end-to-end latency, production memory leaks and pod-scaling instability fixed.
    • Architected the multi-agent research chat on LangGraph across Claude, Gemini and OpenAI, with Qdrant RAG, sandboxed code execution and streaming.
    • Led R&D for a top-5 global pharma client on quantitative systems pharmacology: SBML models, kinetic parameter extraction, paper-vs-simulation checks.
    • Built the extraction eval harness (human ground truth vs model output, LLM-as-judge) and the service template the AI org now starts every service from.
  2. Oct 2022 – Sep 2024

    Giga Tech Limited

    Senior ML EngineerPreviously ML Engineer, Oct 2022 – Dec 2023
    • Built pre-training, data and fine-tuning pipelines for Bangla language models (BERT, XLNet, GPT-2, T5).
    • Led the research behind BanLemma (EMNLP 2023, 96.36% accuracy) and BanNERD (NAACL 2025).
    • Shipped a Bangla text-processing toolkit whose normalisation cut data ambiguity by about 32%, and an internal PyPI server to share it.
  3. Apr 2021 – Aug 2022

    REVE Systems

    Software EngineerPreviously Junior Software Engineer, Apr 2021 – Mar 2022
    • Built streaming and non-streaming Bangla speech recognition; integrating an external language model improved it about 2.2×.
    • Helped build a 10,000-hour Bangla speech corpus with parallel transcriptions for public research use.
    • Added knowledge-distillation continual learning to ASR, keeping 73% coverage on previous test cases while training on new data.

02Kind words

Nafeo AlamLead Software Engineer, DelineateOn LinkedIn

I have worked alongside Shahad at Delineate, where he leads our AI team, and he is one of the calmest, most level headed technical leaders I have worked with. He is young, but remarkably mature in how he approaches problems, people, and leadership.

Two things stand out about Shahad. First, his judgment. He consistently makes the right call, and he does it without drama, even when the problem is ambiguous or the pressure is on. Second, his depth in AI. He understands these systems end to end, from model behavior to production architecture, and that depth shows in the quality of what our team ships.

He is also a strong bridge between product and engineering. He translates in both directions, turning product intent into clear technical direction that engineers can act on, while also surfacing engineering realities in terms that the product side can actually understand. That is a rare skill, and it is a big part of why the team moves as well as it does.

Shahad is easy to talk to, generous with his time, and someone people genuinely want to work with and learn from. Any team would be lucky to have him.

03Research & education

NAACL 2025BanNERD85k-sentence human-annotated dataset; 81.85 macro F1 across 10 entity classes.EMNLP 2023BanLemma96.36% accuracy against a linguist-annotated test set.
Education · 2016–2021B.Sc. in Computer Science & EngineeringRajshahi University of Engineering & Technology. Thesis on deep learning for NLP; competitive programming on the side.

04Open to

Talks

Meetups, conferences, podcasts and university sessions on LLM systems, evals and Bangla NLP.

Guidance

Career and technical guidance for engineers moving into AI and students starting out in NLP research.

Comparing notes

Building LLM systems that have to be accurate? Happy to talk agents, evals and document extraction.

Say hello.Email me LinkedIn