About
I am an AI engineer with almost five years of experience taking models from research to production across LLM agents, NLP and computer vision. At ZS Associates I design agentic AI for pharma R&D and lead development of a production suite of agentic document tools for a global pharmaceutical sponsor. I also research low-resource Bangla NLP, with papers at NAACL 2025 and EMNLP 2023.
I care about systems that are efficient, reliable and explainable, and that keep working when the data changes. Based in Windsor, Ontario, Canada.
Experience
- – Present
Senior Associate Engineer – AI
ZS Associates · Remote from Windsor, Ontario
Agentic AI for pharma R&D in ZS's Drug Development AI practice. I lead development of a production suite of agentic document tools for a global pharmaceutical sponsor.
Highlights (4)
- Development lead, project manager and client contact across the suite. I lead an onshore and offshore team and mentor developers on production APIs and safe agent integration.
- Split an overloaded single-agent classifier into specialised agents, raising internal accuracy by about nine points; production accuracy is close to 90%.
- Ran post-launch improvement cycles with the business users, turning their feedback into extraction fixes that lifted the priority fields from 81–93% to 99–100% accuracy.
- Run regulated releases through specification, UAT and deployment, backed by a live accuracy dashboard and a gold regression dataset in progress.
- –
Senior AI Engineer
micro1 · Remote from Windsor, Ontario
Contract engineer placed with ZS Associates' Drug Development AI practice, building agentic AI systems for a top-10 pharma client.
Highlights (4)
- Designed a tool-based state engine that lets an LLM agent hold its place through long, multi-step conversations where users go back and change answers, with a deterministic engine doing every calculation: about 2x token efficiency, reproducible outputs, and a client demo within a week.
- Generalised a document-analysis pipeline from one table per document to any number within a two-week deadline, bringing previously unsupported documents into scope.
- Made the backend production-grade with async I/O, pooled database and HTTP connections and concurrency locks, for 15–30% higher throughput under concurrent load.
- Split work between onshore and offshore teams by strength, so both work streams ran in parallel and met their deadlines.
- –
Senior AI Engineer
Delineate Inc. · Remote
Multi-agent LLM pipelines, scientific document understanding and the MLOps that keeps them running for pharmaceutical research.
Highlights (10)
- Built a multi-agent data extraction pipeline for PDFs, DOCX and code that beats general-purpose LLMs (Gemini 2.5 Pro, Claude Opus) by 15% on domain-specific extraction such as dosing and covariates, and by 10% on general extraction, evaluated on 150 sources. 10% higher accuracy on tables with more than 100 rows.
- Designed the pipeline to identify multiple fields from a user query, emit structured tables and enforce correct data types.
- Engineered a paper layout system that improved retrieval-augmented generation (RAG) performance by 12%.
- Trained a YOLOv11 subfigure model (global average IoU 0.89) to decompose composite figures and extract captions and legends for visual retrieval.
- Enriched image and table data with section-level metadata, captions, footnotes and summaries to support queries like "breast cancer papers with drug name on the x-axis and dosage on the y-axis".
- Built a production Qdrant vector database over 40M paper PDFs for scalable retrieval.
- Designed a bounding-box filtering heuristic inside NMS that cut false positives by 34%.
- Worked with pharmaceutical scientists on an automated QC system (unit standardisation, column normalisation, biomarker validation, dosing range checks) that raised data accuracy to 97%.
- Set up high-throughput inference with NVIDIA TensorRT and Triton Inference Server; moved long-running work to Celery and RabbitMQ; autoscaled workers on Kubernetes with KEDA from queue length.
- Enforced a zero-trust layout with private networks for every service, VPC endpoints and service-to-service tokens.
- –
Jr. AI Engineer
Giga Tech Ltd. · Dhaka, Bangladesh
State-of-the-art Bangla NLP for a national language corpus programme, plus the data and model infrastructure around it.
Highlights (7)
- Reached state of the art in low-resource Bangla NLP (NER, POS, QA, lemmatisation), exceeding the 90% KPI for every module.
- Set a new NER state of the art (81.85% macro F1, +6% over prior work; 90.49% on the delivered dataset) and POS state of the art with a hierarchical voting scheme over small-model predictions.
- Delivered state of the art on Bangla QA (SQuAD-bn) with a loss that balances null and non-null answers, +6% over the previous best.
- Co-developed and open-sourced the first production-grade Bangla lemmatizer (96.36% accuracy) and an emotion recognition system.
- Built GPT-4o inference pipelines for NER and coreference using ReAct prompting that matched fine-tuned models.
- Fixed severe class imbalance in sequence tagging with sentence resampling and Dice, Focal and CurricularFace losses.
- Architected model and data versioning with DVC and MLflow.
- –
Research Assistant
Qatar Computing Research Institute · Remote
Highlights (2)
- Pretrained a HuBERT model on Bangla ASR data for a joint speech and speaker recognition pipeline using SpeechBrain.
- Enriched open Bangla ASR datasets with scripted audio and corrected annotations.
- –
Undergraduate Teaching Assistant
BRAC University · Dhaka, Bangladesh
Highlights (2)
- Supported coding assignments, lab classes and script checking.
- Helped students run research projects and submit papers to conferences.
Education
- –
BSc in Computer Science & Engineering
BRAC University · Dhaka, Bangladesh · CGPA 3.99 / 4.00
- 100% merit scholarship for 8 consecutive semesters.
- Dean's Prestigious List for academic and co-curricular achievement.
- –
HSC
Notre Dame College · Dhaka, Bangladesh · GPA 5.00 / 5.00
Skills
- LLM & agentic systems
- Claude (API, Agent Skills, Claude Code)Multi-agent pipelinesTool-using agentsContext engineeringModel Context Protocol (MCP)RAGLangChainLangGraphEvaluation harnessesQdrant
- Machine learning
- PyTorchHugging Face Transformersscikit-learnYOLOSpeechBrainWeights & BiasesMLflowDVC
- AWS
- SageMaker AI (Pipelines, Model Registry)LambdaStep FunctionsGlueAthenaLake FormationEventBridgeCloudWatchDynamoDBECS FargateAPI GatewayBedrockQuickSight
- MLOps & infrastructure
- NVIDIA TritonTensorRTDockerKubernetesKEDATerraformCeleryRabbitMQ
- Data platforms
- PythonpandasDataikuApache SparkApache IcebergSnowflakeDatabricksPostgreSQLMongoDB
- Backend & tooling
- FastAPIasyncioDjangoStreamlitNode.jsGitBashLinux
Awards
1st place, BRAC University Inter-University Programming Contest
First among 123 teams from 25 universities.
Certificate ↗- –
Certifications
Claude Certified Architect – Foundations
Anthropic's foundation-level certification for architecting production solutions with Claude.
Certificate ↗AWS Machine Learning Foundations
Preparing, building, training and deploying ML models with Amazon SageMaker and AWS AI services.
Certificate ↗Introduction to Data Science in Python
Coursera (University of Michigan)
Distributions, sampling, hypothesis testing and dataframe manipulation with pandas.
Certificate ↗
Open-source contributions
Fix incorrect token prediction distribution in the Flair sequence tagger
Flair is a framework for state-of-the-art NLP embeddings and sequence models. Fixed a bug in _all_scores_for_token() that produced wrong per-token prediction distributions in sequence classification.