Ovi Poddar / Software engineer
Applied LLMs, agentic systems & trustworthy AI.
I'm Ovi, a software engineer at IQVIA and a BUET graduate with research experience in NLP. My work connects enterprise LLM assistants with hands-on projects in retrieval, agent workflows, and model fine-tuning.
Research
My interests connect applied LLMs and agentic systems with NLP, software engineering, and reliable evaluation.
A study of automated code-review feedback classification using natural language and code representations to support software-engineering analytics.
Contributed to a collaborative PRISMA 2020 systematic review of fairness and bias in histopathology AI, synthesizing 78 studies published between 2017 and 2026.
Work included an eight-category bias taxonomy, analysis of fairness evaluation and debiasing practices, and a proposed reporting checklist and benchmark specification.
With Dr Raktim Kumar Mondol (UNSW) and Shahana Akter · March–May 2026
Read draftExplored code-review comment classification with CodeBERT embeddings and AST/change-based metrics at the Bangladesh University of Engineering and Technology.
Supervised by Dr Anindya Iqbal · February 2022–April 2023
Tool-augmented reasoning, retrieval, agent behavior, and evaluation.
Language and code representations, program analysis, and developer assistance.
Reliability, fairness, bias evaluation, and external validation.
Selected work
From model comparison and fine-tuning to retrieval pipelines and stateful agents. Explore the problem, implementation, and outcome behind each project.
Fine-tuning, agent workflows, and retrieval systems
01 / Model engineering
LLM Engineering Capstone · 2026
Predicting a product's price from its description is a practical way to compare language representations, model families, and adaptation strategies. I built a pipeline from Amazon product-data curation through classical baselines, neural models, API-based LLMs, and QLoRA fine-tuning.
Selected results reported in the project repository. Lower is better.
LLM-assisted text preparation, reusable inference and evaluation, and 4-bit QLoRA adaptation with PEFT. Experiments span a 20K-sample lite configuration and a full 800K dataset.
QLoRA recorded the lowest reported MAE in this project's comparison. Fine-tuning was not uniformly better: GPT-4.1-nano fine-tuning recorded $75.91. These results describe this task and experimental setup.
02 / Agent systems
Designed and implemented a stateful agentic workflow using LangGraph, with specialized nodes for conversational reasoning, web search, and news aggregation. Implemented state management, conditional routing, caching, retries, and tool integration to support reliable multi-step agent execution. Built with Streamlit for a unified interface and Tavily API for real-time search.
Workflow overview based on the documented architecture.
Separate graph nodes, shared conversation state, tool integration, cached graph compilation, and retries with exponential backoff.
A working chat, search, and news application with documented demos. The repository does not report a controlled task-success or latency benchmark.
03 / Retrieval engineering
Built a retrieval-augmented news intelligence system using OpenAI, LangChain, and FAISS. Users load article URLs or text files into a semantic index and ask context-aware questions with cited responses. Designed an end-to-end RAG pipeline with Streamlit UI for efficient research workflows.
Pipeline overview based on the project documentation.
Source ingestion, semantic retrieval, a reusable local FAISS index, and a Streamlit question-answering interface.
A documented application and video demo for exploring indexed articles. Citations support source inspection; the project does not establish answer faithfulness through a formal benchmark.
Developed a multi-agent system using Model Context Protocol (MCP) and Google Gemini, orchestrated with LangGraph. Built and connected custom MCP servers—Weather (Go) and Calculator (Python)—through Gemini's tool-use framework to answer weather and calculation queries via tool-augmented agent reasoning.
Designed a multi-stage agentic pipeline for automated blog generation. Specialized agents handle topic generation, web research, outline planning, and article drafting, coordinated by a central workflow controller for structured reasoning and modular content generation.
Developed an AI-assisted personal finance application that tracks expenses, analyzes spending patterns, and generates intelligent insights. Applies AI-driven analysis to provide personalized budgeting recommendations and expense optimization.
Structured implementation work — weekly labs across LLM engineering and agentic systems
An implementation-focused eight-week exploration of modern LLM engineering — from LLM applications and multi-model systems through Hugging Face/Transformers, RAG, vector databases, fine-tuning (LoRA/QLoRA), deployment, and agentic AI. Repository organized by week for inspectable, hands-on progress.
Structured labs covering OpenAI Agents SDK, LangChain/LangGraph, CrewAI, and Model Context Protocol (MCP). Built and exercised agent workflows spanning foundations, tool integration, multi-agent orchestration, and custom MCP servers.
Full-stack web applications and embedded systems
Full-stack MERN application where users can discover local businesses, create business profiles, leave reviews, provide star ratings, and upload photos.
Instagram-like social networking platform supporting account authentication, photo sharing, following, likes, comments, notifications, and chat.
Digital alarm clock system using ATMEGA32 microcontroller integrated with an SD-Card module for data storage and real-time clock functionality.
Open source
Work on evaluation correctness, training reliability, retrieval systems, and parameter-efficient fine-tuning.
· 6 open pull requests · 1 open issue
Temporal retrieval
and graph reliability
A 21-case Neo4j benchmark measured when historical facts were returned, missed, or mixed with facts from the wrong period. A separate patch lets Neo4j save entities with optional embeddings.
Submitted documentation and a regression test for point-in-time retrieval, including queries at the boundary between two facts.
Independently reproduced two retrieval failures and validated existing patches for combined date filters and full-text candidate limits. The second patch recovered the valid fact at rank one behind 30 superseded candidates.
Recoverable
document deletion
Partially deleted documents could retain a healthy-looking status. The submitted change exposes deletion-in-progress state and keeps interrupted work visible for retry.
Problem: interrupted cleanup can leave document status inconsistent with stored data.
Change: track DELETING state, isolate it from ingestion, preserve it during scans, and expose it in the interface.
Evidence: focused deletion lifecycle and frontend tests described in the PR.
Reward validation
for agent training
A value valid as a Python float can become infinity in float32. The submitted change rejects non-finite rewards and out-of-range training values before conversion.
Cause: a finite value such as 1e40 exceeds float32's finite range.
Change: validate rewards and fallbacks, with errors identifying the rollout. Preserve large finite validation metrics.
Evidence: boundary cases and a local event-API-to-training-tensor reproduction.
More reliable
answer evaluation
The evaluator could extract A from “Answer choice: D.” The submitted fix prioritizes explicit answer formats and narrows the fallback to a unique standalone label.
Cause: an unconditional regex selected the first capital A/B/C/D anywhere in the response.
Change: recognize bare labels, retain structured patterns, and use a unique standalone-letter fallback.
Evidence: failing baseline tests, passing unit and CLI tests, and open-ended QA controls.
Layer selection
in adapter merging
Weighted merging can inherit the first adapter's exclusion filters and drop layers from another source. The submitted fix preserves the layers each source actually adapts.
A PEFT maintainer approved work on the fix.
Tests cover both source orders, list and regex exclusions, and saving and reloading the combined adapter. Existing LoRA and IA³ checkpoints produced identical outputs before and after the change in three CPU dtypes.
Industry experience
Enterprise AI assistants, domain-aware evaluation, and full-stack software systems.
Enterprise LLM assistants · Life-sciences analytics
I translate domain requirements, client feedback, and data semantics into prompts, agent instructions, and behavioral constraints.
Full-stack engineering · E-learning
Contributed to an education platform using ASP.NET Core, Angular, and SQL Server, building application modules, improving performance, and resolving reliability issues.
Along the way
Research, engineering, and the steps in between.
Latest first Scroll for earlier updates
Verified the LoRA and IA³ fix in eight CPU regression cases, including saving and reloading the combined adapter.
Ran 21 retrieval cases on Neo4j and reproduced two ways a search can miss a fact that was valid at the requested time.
Fixing answer extraction that could both reject correct responses and reward incorrect ones.
A CPU reproduction of source-order-dependent weighted merging in LoRA and IA³.
Checking reward finiteness before conversion to training tensors.
Tracking deletion state so partially deleted documents remain visible and retryable.
Collaborative systematic review covering 78 studies, bias sources, evaluation practices, and reporting recommendations.
Started as a Software Development Engineer, working on enterprise software and, subsequently, LLM-powered assistants.
“Towards Automated Classification of Code Review Feedback to Support Analytics” received full acceptance.
The revised manuscript and response to reviewers were submitted following conditional acceptance.
Started as a Software Engineer I, contributing to an e-learning platform with ASP.NET Core, Angular, and SQL Server.
Explored NLP and code representations for code-review feedback classification, supervised by Dr Anindya Iqbal.
About
A software engineer with a research background in NLP and software engineering.
I work at the intersection of LLM behavior, domain requirements, and evaluation. Alongside my work at IQVIA, I build projects in RAG, parameter-efficient fine-tuning, and agentic systems.
My research interests span applied LLMs, agentic AI, NLP, and trustworthy AI. I am interested in PhD opportunities that connect these areas with practical, reliable systems.
Download CVCapabilities
Recognition
Awarded the Dean's List distinction at BUET for maintaining a CGPA of 3.75 or higher across all semesters of Level-2.
Solved 2500+ problems on Codeforces, LeetCode, and other platforms, demonstrating expertise in algorithms, data structures, and optimized problem-solving.
Recognized for delivering higher-than-expected contributions, taking ownership in the absence of a tech lead, and ensuring stability and timely feature delivery.
Awarded the Talent Pool Scholarship in both Secondary and Higher Secondary Examinations in Bangladesh for outstanding academic performance.
Contact
For research collaborations, PhD opportunities, and conversations about applied AI.