AI / ML Engineer

Shamim Ahamed

Compliant, evaluated, production AI for healthcare and other high-trust domains, built so it's measurably reliable.

Shamim Ahamed
Work authorisation & relocation
Bangladeshi
Open to relocation to Germany or Ireland
Germany
anabin H+ · degree 'Entspricht' · EU Blue Card §18g AufenthG
Ireland
Eligible for the Critical Skills Employment Permit
Ready
Documents can be filed the week a contract is signed

Production AI you can
actually trust1.

I build and run a clinical LLM platform in daily production for a US healthcare company, and I publish peer-reviewed research on how such systems are evaluated. I focus on healthcare and other high-trust, regulated domains.

Daily production
Clinical LLM platform in daily production use at a US clinic
9
Peer-reviewed publications, 3 first-author (IEEE · AAAI · Elsevier · BMC)
Gold
Kaggle Gold Medal, top 2%, Google-sponsored, plus a Silver
SuSastho.AI
Shipped multimodal medical copilot, now an Elsevier paper
01 The short version

For over two and a half years I've built and run HIPAA-aware, tool-using agent AI for a US healthcare company, the kind of system where a wrong answer isn't a bug, it's a liability. That work taught me the thing most AI teams skip: shipping the model is the easy 20%. Proving it's safe and grounded is the other 80%.

So that's what I build. I both build production clinical AI and publish on how to evaluate it: I'm a published researcher on LLM evaluation and health-triage benchmarking, so I don't just claim my systems are reliable, I measure how reliable they are. Compliance and high-trust domains scare most generalists off. For me they're the whole point.

"Honesty is a feature. If a system can't show its confidence, its sources, and its cost, it isn't production-ready."

02 What I build
01

Compliant clinical & healthcare AI

Agentic systems for high-trust domains, safety gates, audit trails, and a real story for where data lives and what leaves the boundary.

02

Evaluated, observable RAG

Retrieval systems that cite their sources and are measured for faithfulness, not vibes, so you know when to trust the answer and when not to.

03

LLM evaluation & reliability frameworks

The eval harness most teams skip: gold sets, distributional reporting, and a meta-eval that scores the evaluator itself.

04

Cost & latency optimisation

Routing, caching, and cascades that cut inference spend without quietly trading away quality, with the Pareto frontier to prove the tradeoff.

llm-cost-router
03 Professional experience

AI Engineer

FRQ Technologies
Massachusetts, USA (remote)
Dec 2023 – Present
  • Build and run the clinical AI platform of a US healthcare technology company: a multi-stage tool-using agent pipeline covering patient support, medical Q&A, and care-workflow automation, in daily production use at a US clinic. Every response reaches a real patient over real PHI.
  • Own the retrieval layer grounding every answer in curated clinical sources, indexed in a production vector store with chunking and embeddings tuned for clinical language. Questions outside approved clinical scope are refused, not answered.
  • Own the platform end to end on Azure and AWS: FastAPI in Docker, PHI-bearing PostgreSQL, CI/CD, and LLM inference on Azure OpenAI under an agreement covering PHI (chosen over a direct API for exactly that reason).
HIPAA-awareRAGAgent pipelineAzure OpenAIFastAPIAWS

Research Engineer, then Research Associate

AIMS Lab, United International University
Dhaka, Bangladesh
Nov 2023 – Present
  • Built and deployed SuSastho.AI, an evidence-based multimodal medical copilot (Llama-3 8B fine-tuned on 8,614 clinician-authored examples, RAG over a clinically validated knowledge base). In a clinician-scored pilot over 428 questions it produced 26.9 points fewer incorrect answers and 16.1 points higher correctness than the baseline Llama-3 8B. First-author paper, Informatics in Medicine Unlocked (Elsevier, 2025).
  • Designed the human evaluation protocol behind those numbers (two independent clinical evaluators, third-party adjudication, correct refusal scored as correct), and now lead the first-author follow-up on LLM-as-a-judge evaluation of medical chatbots (under review). Co-author of an ML/LLM health-triage benchmark, IEEE ICDM Workshops 2025.
Fine-tuningRAGLLM evaluationClinical NLP
04 Skills
LLM systems
RAG (chunking, embeddings, vector search), agent pipelines and tool use, fine-tuning (Llama-3 8B, instruction tuning), quantisation, prompt design. Python, PyTorch, Hugging Face Transformers, LangChain, Azure OpenAI.
LLM evaluation
Clinician-adjudicated gold sets, inter-rater protocols, LLM-as-a-judge, refusal and out-of-scope safety testing.
Backend & cloud
FastAPI, async Python, REST APIs, PostgreSQL, Redis, Docker, CI/CD, Azure, AWS (EC2, S3, VPC), Linux, Git.
Regulated-domain AI
PHI handling and HIPAA-aware design, clinical NLP, safety guardrails and scope refusal, human-in-the-loop clinical review.
Languages
English (full professional proficiency, daily work with a US team since 2023); German (A1, in progress).
Education
  1. 01
    B.Sc. Eng. in Computer Science & Engineering
    Khulna University of Engineering & Technology (KUET) · Khulna, Bangladesh · 2018 – 2023
05 Selected work

Four projects spanning production clinical AI and safety-first agent engineering. One is shipped and peer-reviewed; the rest are built, CI-evaluated, and open-sourced, each with real numbers instead of vibes.

01 Shipped · Published

SuSastho.AI

An evidence-based multimodal medical copilot for adolescent health in Bangladesh, built on Llama-3, RAG, and clinical chain-of-thought reasoning.

Peer-reviewed
Elsevier · 2025
Vs baseline
26.9 pts fewer incorrect answers
Pilot
428-question clinician pilot
Healthcare RAG Multimodal LLM
Read the paper GitHub
SuSastho.AI chat interface answering medical questions in Bengali and English
Any chest pain or pressure right now?
Yes, tightness since this morning.
!
Safety gate · escalate
Route to ER, chest-pain red flag
02 Built · CI-evaluated

ScribeIntake

A pre-visit intake agent that interviews a patient, runs a deterministic, code-based safety gate ahead of the LLM, and hands the clinician a cited SOAP summary and triage band.

Signature metric
0 missed escalations on the must-escalate set, CI-gated on every commit
Agentic RAG Eval HIPAA-aware
GitHub
03 Built · CI-evaluated

groundcheck

A lie detector for LLM answers: it decomposes a response into atomic claims and checks each one against the source, sentence by sentence.

Signature metric
Two-tier meta-eval: the detector itself is scored against a frozen gold set
Eval Grounding Hallucination LLM-as-judge
GitHub
groundcheck decomposing an LLM answer into atomic claims and scoring each against the source, 62% faithful
Relay holding a prompt-injection attempt at its deterministic approval gate, awaiting human approval
04 Built · CI-evaluated

Relay

An ops-automation agent that triages a ticket and runs a tool-use loop against a backend, but pauses every state-changing action for human approval, held at a deterministic code gate rather than a prompt.

Signature metric
0 un-approved actions across 88 CI-gated eval runs
Agentic Tool-use Safety Multi-provider
GitHub
06 Endorsement

"Shamim is highly skilled at building systems using Large Language Models. He's always ready to lend a hand and takes feedback constructively. It's been a pleasure to work alongside him, he consistently contributes quality work and fosters a positive, collaborative environment."

CEO, FRQ Technologies, USA
07 Research & recognition

9 peer-reviewed publications, 3 first-author, plus 2 manuscripts under review. Reliability is something I measure and publish, not something I assert.

Selected, peer-reviewed
  1. 01
    SuSastho.AI: A Multimodal Medical Copilot for Adolescents Using Evidence-Based Medicine and LLMs
    First author · Informatics in Medicine Unlocked (Elsevier), 2025 · DOI ↗
  2. 02
    A Benchmark and Optimization Framework for ML and LLM-based Health Triage
    Co-author · IEEE ICDM Workshops 2025 · presented, Washington D.C.
  3. 03
    Artificial Intelligence-Based Risk Assessment Tools for Sexual, Reproductive and Mental Health: A Systematic Review
    Co-author · BMC Medical Informatics and Decision Making, 2025
  4. 04
    Enhancing Access to Health Information for Adolescent Girls in Bangladesh Through a Culturally Responsive AI Chatbot
    First author · Gender and Technology Conference (GTC), 2025
  5. 05
    Integrating AI-Driven Triage into Digital Pharmacy Systems for Rational Antibiotic Use in Low-Resource Settings
    Co-author · AAAI 2026 Bridge on AI for Medicine & Healthcare · Singapore
Under review
  1. 01
    Trustworthy & Scalable Evaluation of Medical Chatbots: Benchmarking LLMs as Automated Judges, and the Role of Human–AI Hybrid Oversight
    First author · Under review · LLM-as-a-judge
See all publications on Google Scholar
Recognition
  • Kaggle GoldTop 2%, March Machine Learning Mania 2022 (Google-sponsored); Silver, Bengali.AI Speech Recognition (26th of 744 teams)
08 Let's talk

Looking for an AI/ML engineer who can ship in a regulated domain?
Let's talk.

I'm open to full-time roles across the EU and ready to relocate. Send a few lines about the role.

shamim.aiexpert@gmail.com  ·  +880 1324 143904