Machine Learning & AI Engineer
Dhaka, Bangladesh
--:-- local
AHAMMADNAFIZ
I build retrieval systems and LLM agents that hold up outside a notebook.
Right now I'm a research assistant at UIU, working on autonomous agents and reading a lot about mechanistic interpretability. Before that I was a junior engineer at AskTuring.ai, building the evaluation and retrieval behind a production RAG system.
Open to ML / AI roles & research collaborations
- Now
- Research Assistant, UIU
- Before
- Jr. Software Engineer, AskTuring.ai
- Focus
- RAG · LLM Agents · Interpretability
- Goal
- PhD in Europe, then a frontier lab
About
I'm a data science student at UIU and an ML engineer. My last role was junior engineer at AskTuring.ai, where I owned the evaluation service and built Deepen Answer, a RAG pipeline that pairs extended-thinking models with an agent that fills its own knowledge gaps.
I tend to learn things by rebuilding them. An autograd engine in NumPy. A memory layer for agents that cut context cost by 22× without dropping accuracy. A research agent that plans, searches, and cites its sources. Lately I keep drifting toward interpretability, trying to work out what these models are actually doing inside.
Next I want a PhD in Europe, and after that a lab working on the frontier.
[ Experience ]
United International University
Dec 2025 — PresentResearch Assistant
Researching autonomous agents and multi-agent systems, and helping with papers on LLM applications, retrieval, and agentic workflow design.
AskTuring.ai
Sep 2025 — May 2026Jr. Software Engineer (former)
Owned the eval service and admin dashboard, automating RAG benchmarks over HotpotQA, MuSiQue, FinQA, and CUAD. Built Deepen Answer: extended-thinking LLMs, an agent that fills its own knowledge gaps, and hybrid BM25 + vector retrieval with RRF.
B.Sc. in Data Science — UIU
2023 — Present[ Recognition ]
Oct 2025
Champion — IC6 University Innovation Hub
BDT 60,000 pre-seed after a 15-week program.
Jan 2025
Best Prototype — Blockchain Olympiad
Team Auctra, UIU CSE Fest. BDT 20,000 prize.
Jun 2024
Champion — CSE Project Show (Predicta)
End-to-end data analysis & ML platform.
2024
1st Runner-Up — Data Innovators Challenge
Sales analytics track.
Skills
- Languages
- PythonC++SQLBashCUDA
- ML / DL
- PyTorchscikit-learnXGBoostCatBoostNumPyPandasOpenCVMatplotlib
- LLM & GenAI
- LangChainLlamaIndexLangGraphHF TransformersCrewAIAutoGenvLLMOpenAI APIAnthropic API
- Agent Stack
- ReActTool UseFunction CallingMulti-AgentSupervisor / SwarmLLM MemoryRAGLLM Eval
- Fine-tuning
- LoRAQLoRAPEFTSFTRLHFInstruction Tuning
- Inference
- vLLMONNXTensorRTBentoMLTorchServeQuantization
- Backend / MLOps
- FastAPIDockerKubernetesMLflowW&BDVCGitHub Actions
- Databases
- PostgreSQLpgvectorMySQLChromaFAISSPineconeWeaviate
Domains — RAG · LLM Evaluation · AI Agents · NLP · Computer Vision · MLOps · Mechanistic Interpretability
Work
- 01
fizzlabai
Deep Research Agent
A solo deep-research agent running a deterministic 8-phase pipeline (plan, search, read, synth, verify, re-plan, polish, close) over the open web, producing cited Markdown with a full execution trace.
LangGraphFastAPIpgvectorOTel
- 02
Engram
Persistent Memory for AI Agents
A memory library on PostgreSQL + pgvector combining hybrid search (BM25 + vector + RRF), graph traversal, and durable task ledgers. Cut LLM context cost up to 22× at 90%+ on LongMemEval, LoCoMo, and BEAM.
pgvectorHybrid SearchInfra
- 03
Personal Knowledge Assistant
Multi-Agent RAG System
A multi-agent RAG system with confidence-based routing, knowledge-strip decomposition, and automatic web-search fallback for low-confidence queries.
Multi-AgentRAGRouting
- 04
Analyzia
AI-Powered CSV Analysis
A LangChain agent that turns plain-English questions over CSV data into pandas analysis and Plotly visualizations, with confidence-scored answers and conversation memory.
LangChainAgentsPlotly
- 05
SmartPrice Intelligence
Smartphone Market Analytics
Scraped 10,000+ listings across 37 brands from GSMArena under adaptive rate limiting, then trained CatBoost / XGBoost on 50+ engineered spec features (R² = 0.772, 74.4% price-segment accuracy).
CatBoostXGBoostScraping
- 06
FizTorch
Deep Learning Framework
A NumPy-based deep-learning framework from scratch: autograd engine, core layers, and an Adam optimizer, validated on MNIST and California Housing.
NumPyAutogradFrom Scratch
Contact
If you're hiring for ML or AI work, want to talk research, or just have a good problem worth nerding out over, my inbox is open.
ahammadnafiz86@gmail.com