Ishrat Jahan Ananya

ML Engineer specializing in production information retrieval, multimodal reranking, and LLM evaluation.

Paris, France · nahajtarshi@gmail.com

Summary

ML engineer specializing in information retrieval and multimodal systems: fine-tuning rerankers, building vision-language pipelines, and shipping them into production.

I did my Master's in CS (AI focus) at Illinois Institute of Technology, Chicago, and most recently worked at LightOn in Paris on neural reranking and retrieval evaluation, where I trained and released the LightOn-rerank model family and first-authored the technical report behind it. I care about the gap between a paper and a working system, and closing it fast.

Role
ML Engineer
Location
Paris, France
Email
nahajtarshi@gmail.com
Availability
Open to full-time ML Engineer, Applied Scientist, and Information Retrieval roles across Europe and North America. In-person, hybrid, or remote.

What I bring

I work at the intersection of IR research and production ML, fine-tuning retrieval systems, building multimodal pipelines, and shipping them as real products.

What I'm looking for

Full-time IR or ML Engineer / Applied roles where I can own the full stack from model to system.

Looking for the next chapter · Full-time roles, Europe & North America. In-person, hybrid, or remote.

Experience

ML Engineer

LightOn · Paris, France · Jan 2026 – Jul 2026

  • Trained and shipped LightOn-rerank: six public checkpoints (listwise and pointwise at 0.8B / 2B / 4B) built on a single Qwen3.5 LoRA adapter that scores both text passages and document-page images. First author on the accompanying technical report.
  • 2B listwise reaches 62.66 nDCG@10 on ViDoRe V3 (+7.06 over the ColQwen2.5 first stage), ahead of Qwen3-VL-Reranker-2B and jina-reranker-m0 in the same parameter class.
  • Pointwise checkpoints integrated into the sentence-transformers library; MTEB integration PR in review.
  • Fine-tuned Qwen as a text cross-encoder reranker via rsLoRA and shipped it into Paradigm, LightOn's production platform, giving 7.4% mean nDCG@5 over the previous production baseline on BEIR.
  • Designed and shipped a unified evaluation infrastructure spanning retrieval reranking, VQA, code generation, tool routing, and parsing, adopted as the team standard across all model and feature work.
  • Built the retrieval and eval harness for the team's omnimodal retrieval model, now used as the standard evaluation pipeline.
  • Constructed a custom training dataset for listwise reranking via knowledge distillation from larger teacher models, enabling the Plackett-Luce loss regime experiments.
  • Distributed training across H100 GPUs with FSDP; managed the full pipeline from data preparation through training and eval.

AI Software Engineer

Wordpower AI · Chicago, US · Jan 2024 – Dec 2025

  • Built an agentic NL-to-SQL system over 100K records (schema-aware retrieval → query drafting → execution → self-correction loop), reaching 91% execution accuracy end-to-end.
  • Engineered a production RAG pipeline using dense retrieval on AWS (Lambda, S3), serving real-time LLM responses to external clients with sub-2.5s end-to-end latency.
  • Designed and scaled a vector retrieval system over 15K+ documents in Pinecone, achieving 30% retrieval precision improvement through embedding tuning and cross-encoder reranking.
  • Built a multi-agent IR pipeline for B2B news curation using dense retrieval, LLM summarization, and editorial feedback loops, reducing client article research time by ~60%.

Research Assistant

NLP Lab, Illinois Tech · Sep 2023 – Dec 2023

  • Investigated the use of LLMs for text domain adaptation leveraging targeted prompt tuning
  • Designed LLM based pipeline for authorship obfuscation task

Research Assistant

Scalable Computing Software Lab, Illinois Tech · Feb 2023 – Aug 2023

  • Developed an intelligent cache eviction algorithm for Deep Memory and Storage Hierarchy System
  • Implemented a profiling tool for analysis and assessment of I/O performance

Deep Learning Engineer

CPSD Technologies (XR Interactive) · Feb 2021 – Jun 2022

  • Achieved 96% accuracy by integrating ArcFace into facial recognition models.
  • Trained and deployed MTCNN for Bengali face detection and downstream verification.
  • Compressed face recognition models via NCNN for real-time edge deployment.
  • Built internal analytics dashboards to monitor AI system outputs and user behavior patterns for ops and product teams.

Research, publications and model releases

One Adapter, Both Modalities

Technical Report + Model Release · LightOn · 2026

First-authored the technical report behind LightOn-rerank and trained the model family: a single LoRA adapter on Qwen3.5 that scores both text passages and document-page images against a query, with no separate text and vision heads.

  • Released six checkpoints: listwise and pointwise variants at 0.8B, 2B, and 4B.
  • 2B listwise reaches 62.66 nDCG@10 on ViDoRe V3, +7.06 over the ColQwen2.5 first stage, ahead of Qwen3-VL-Reranker-2B (59.18) and jina-reranker-m0 (59.40) at the same scale.
  • Showed listwise scaling holds where pointwise plateaus: +10.0 nDCG from 0.8B→2B and +2.0 from 2B→4B listwise, versus −0.1 for pointwise at 2B→4B.
  • Cross-lingual transfer from English-only training: 61.31 nDCG@10 on French ViDoRe V3 queries against 64.01 on English.
  • Profiled inference to find ViT encoding, not decoding, dominates cost (~46% of per-window GPU time at 2B listwise); reranking the top-20 instead of top-100 keeps 85% of the quality gain at 5.4× fewer windows.
  • Pointwise models are integrated into the sentence-transformers library; MTEB integration PR currently in review.

Qwen3.5 LoRA Listwise Reranking ViDoRe V3 BEIR sentence-transformers vLLM

Acoustic Scene Classification

Conference Paper · MICAI 2021 · Mexico

Explored CNN-based acoustic scene modeling for environmental sound understanding.

CNNs PyTorch OpenCV NumPy

Comparative Analysis of Six Programming Languages

Conference Paper · ICCIT 2021

Published at the 24th International Conference on Computer and Information Technology. Empirically evaluated six mainstream programming languages on readability, writability, and reliability, combining qualitative and quantitative metrics for software engineering education.

Research Java Python C++ Technical Analysis

Projects

WanderLust 3rd · OpenBMB Prize

Status: Live

Third place ($1,000) in the OpenBMB Prize at Hugging Face's 'Build Small' hackathon, Backyard AI track. Reimagines navigation around discovery over speed: takes a start, destination, a free-text 'vibe', and an adventurousness level, then returns a walking/biking route that detours through interesting places within a travel-time budget. Models routing as an Orienteering Problem solved with a budgeted greedy heuristic over a submodular reward (diminishing returns per category for diversity), runs multi-source Dijkstra over OpenStreetMap graphs, cutting latency from 8–14s to roughly one second, and gates narration behind a fail-closed grounding verifier that validates every place name against allowed waypoints. Runs entirely inside the Space with no external API calls at request time, with nine cities preloaded at boot.

MiniCPM5-1B BGE-small OSMnx SciPy ZeroGPU Gradio

VigilanceAI

Status: Development

World model anomaly detection for video: detects anomalies by measuring when V-JEPA 2 is surprised. Uses spatial variance of per-patch prediction error in latent space as the anomaly signal. Indexes video segments in Qdrant for temporal RAG: retrieve moments across a library by physical surprise pattern.

Python PyTorch V-JEPA 2 HuggingFace Qdrant Gradio torchcodec

TrademarkVista

Status: Live

Built a RAG-based multilingual AI agent using SmolLM and LangChain for USPTO trademark search. Integrated GraphQL APIs and conversational retrieval logic.

GraphQL RAG PyTorch Docker AWS Vercel

Prospect

Status: Live

Autonomous AI job hunter for Twitter/X. Scrapes hiring tweets, scores them against your profile using a local or cloud LLM, and surfaces the best leads in a Streamlit review dashboard. Supports daemon mode, query expansion, and a full outreach pipeline.

Python Playwright Streamlit SQLite Ollama SQLAlchemy

Navigational Aid for the Visually Impaired

Status: Live

Developed a semantic segmentation-based voice navigation system to assist with object detection and scene awareness.

Deep Learning Python DeepLabV3 Java PyTorch OpenCV Android

Bangle Chat AI

Status: Development

Built a Bengali-language RAG agent with SmolLM for low-resource document retrieval.

RAG OpenAI Flask PostgreSQL React

Annotated-ViT-Paper

Status: Live

Implemented Vision Transformer Architecture from scratch for image classification on CIFAR 10 dataset.

Python Transformer PyTorch LaTeX

GEC-with-SFT-SmallLM

Status: Live

Implemented Supervised Fine-Tuning (SFT) of the SmolLM-135M model for the Grammatical Error Correction (GEC) task on CoEdIT dataset

SFT LLM PyTorch Finetuning Hugging Face Transformers

Skills

Information Retrieval & Ranking

Cross-encoder Reranking, Listwise Reranking, Sliding-Window Decoding, Dense Retrieval, BM25, Hard Negative Mining, Listwise / Pointwise Loss, Plackett-Luce, Knowledge Distillation, BEIR / MIRACL / ViDoRe V3, nDCG / MRR / MAP, sentence-transformers, MTEB, Embedding Tuning, RAG Pipelines, Vision RAG, Document Search, VQA

Multimodal & Vision Systems

VLM Fine-tuning, Vision-Document Retrieval, ViDoRe / ColPali, TTVT Modality Interleaving, World Models, CNNs / YOLO, Gradient Conflict Mitigation, Anomaly Detection

LLM Training & Optimization

LoRA / rsLoRA / PEFT, FSDP Distributed Training, Knowledge Distillation (Soft+Hard), SFT / DPO / RLHF, Ablation-driven Experimentation, vLLM Inference, Quantization, Multi-GPU Scaling

Production AI Engineering

End-to-end RAG Systems, Agentic Pipelines, Vector Databases (Pinecone), NL-to-SQL, Serverless Deployment (AWS Lambda), Docker / FastAPI, Multi-task Eval Infrastructure, SaaS API Architecture

How I work

Education

Languages

English (Fluent), French (A2, actively learning), Bengali (Native)

Community and volunteer work

Outside the lab

Languages fascinate me, not just as communication tools but as proof that different communities have built entirely different frameworks for experiencing the world. Words like ঢং (dhong) or আদর (ador) carry something English simply has no container for. That gap is where culture lives. It's a big part of why I care about multilingual AI.

I'm drawn to unconventional homes. The idea that a living space should be shaped around the person who inhabits it, not the other way around, resonates with how I think about building software too.

I have a quiet obsession with symmetry in nature, design, and systems. Near-symmetry is often more interesting than the perfect kind; it implies intention just barely interrupted.

Contact