AI/ML Engineer at Varosync, building the drug-development ML stack from scratch: knowledge graphs, molecular-similarity models, and the pipeline behind them. MS AI at Northeastern Khoury (4.0). I ship interpretable ML for problems where being wrong has consequences.
“What stands in the way becomes the way.”— Marcus Aurelius
My AI Digital Twin is live on HuggingFace. Ask about my work, have it write a cover letter, assess my fit for a role, or pitch how I'd help your team. First load takes about 20 seconds while the Space wakes.
I build ML systems for problems where being wrong has consequences. Right now that means Varosync, a computational drug-development startup where I'm the only person on the ML stack, building it from scratch: the knowledge graph, the molecular-similarity models, and the pipeline behind them, on production infrastructure across AWS, Nebius, and Mithril.
Before this: medical imaging research at CMATER Lab (Jadavpur), a published Springer paper on traffic sign detection, an IEEE paper with a patent filed on lightweight intrusion detection, and a PhD-level mechanistic interpretability seminar at Khoury. 4.0 GPA across both my BS and MS. This summer I'm teaching the ML lab for Northeastern's AIDE program, a Sloan-funded residency bringing philosophy PhDs into hands-on ML, fairness, and interpretability.
I haven't found the problem yet that made me want to slow down.
Six years, summarized: four internships, two published papers, one patent filed, three teaching posts, forty-plus students taught, one selective AI-safety cohort, and a 4.0 in both degrees. Still going.
Northeastern University · Khoury College of Computer Sciences · Boston, MA
Institute of Engineering & Management · Kolkata, India
The only person on Varosync's ML stack, building it from scratch and reporting directly to the CEO and CTO. I own it end to end: the knowledge graph, the molecular-similarity models, and the pipeline through to the live query path.
Teaching Co-Lead, Machine Learning Lab
Northeastern University · Part-time alongside Varosync
Nine-week summer intensive led by Kathleen (Katie) Creel, funded by the Alfred P. Sloan Foundation, bringing together ~10-12 philosophy PhD students from institutions across the country for training in ML, AI ethics, and metascience. Designing and leading a 12-session hands-on ML lab built from scratch: covering Python, fairness, differential privacy, neural networks by hand, transformers, and mechanistic interpretability. Adding new metascience and agent-based modeling sessions this year.
Teaching Assistant, NLP
Khoury College, Boston, MA
Taught the PyTorch and deep-learning labs for Professor Amir Tahmasebi's graduate NLP course, from fundamentals through transformer architectures.
Developed custom ML pipelines for cell segmentation in histopathological images under Dr. Kaushiki Roy and Dr. Debotosh Bhattacharjee. Implemented preprocessing techniques using color quantization and DBSCAN clustering for densely packed cell detection. Guided junior researchers in deep learning for medical imaging.
Contributed to developing an Indian Sign Language dataset under Dr. Arnab Kumar Maji. Defined quality criteria and annotation standards by analyzing existing datasets including Microsoft ASL, MNIST, and Static ISL to establish best practices.
Facilitated IT system migration from local servers to Oracle cloud infrastructure. Supported IT operations for a 1000+ room property.
Taught mathematics and computer science to 10 underprivileged students over seven years. Developed custom learning materials and structured lesson plans that broke complex topics into manageable concepts.
CS 7180 · Byron Wallace · Khoury College
Falsified a standard assumption in mechanistic interpretability: separator tokens score 1.000 LDA accuracy across all 32 layers of Llama-3-8B but cause 0% prediction flips under causal patching, proving probing accuracy and causal relevance are independent properties. Built a three-part pipeline (LDA + gradient attribution + causal patching), applied to LLM resume screening where name-swapped resumes flip at 95.8%.
Live on HuggingFace · Chat + generate
A RAG twin with four modes: chat about my work, a tailored cover letter, an honest job-fit read, or a pitch for how I'd help your team, all grounded in my own docs. all-MiniLM-L6-v2 embeddings, in-memory cosine retrieval, OpenRouter gpt-oss-120b (Llama 3.1 8B fallback). 2-5s responses.
MCP · OpenAI Agents SDK
4 AI traders (Warren, George, Ray, Cathie) with 6 MCP servers, 44 tools, each managing $10k portfolios.
CS 5100 · Khoury College
Classical ML (Random Forest: 0.94 AUC) vs embeddings (TransE: 0.61 AUC) on BioRED knowledge graph.
Independent Research · ICLR 2023 Reproduction
Mechanistic interpretability study reproducing Nanda et al. (ICLR 2023). Discovered that a 1-layer transformer trained on modular addition internally learns discrete Fourier transform, an algorithm no human would guess from I/O behavior alone.
5 Complete Projects
Engineering Team, Stock Picker with hierarchical memory, Multi-LLM Debate (GPT-4o vs Claude), Financial Researcher, and Docker-based code execution.
Trained lightweight linear probes on each layer's activations to map where token-level signals — NER, POS, chunking — actually live inside a transformer, then used that map to guide layer-specific quantization. Keeping precision only in the layers that carry the signal holds 99–100% accuracy at a 5-bit average, where uniform quantization collapses to 16–41% on unseen datasets.
Read on LessWrongIntegrated YOLOv8 with a custom CNN filtering layer to minimize false positives in cluttered Indian driving environments. Published in Springer's "Information and Communication Technologies" series at AISC 2024.
Designed custom ML pipeline combining classical computer vision and deep learning under Dr. Kaushiki Roy at Jadavpur University. Implemented DBSCAN clustering for detecting densely packed cellular structures.
A 80K-parameter hybrid architecture combining a statistical DNN branch with a pseudo-GNN (MLP) branch fused via learnable adaptive weights. Achieves 99.64% accuracy on CIC-IDS-2017, outperforming Transformer-based systems with 1,250x fewer parameters. Enables real-time inference at over 16,000 flows/second on edge hardware.
View on IEEE XploreAccepted into BlueDot's competitive technical AI safety cohort, covering interpretability, alignment research, and frontier safety methods.
Deep Learning and AI
ML Frameworks and Data Science
LLM and GenAI Engineering
Programming and Algorithms
Specialized Skills
Competitive Programming
I'm finishing my MS in AI at Northeastern (May 2027) while building drug-trial-prediction models at Varosync, Inc. Now I'm looking for what's next: Spring 2027 roles starting January, in ML engineering, AI research, or applied AI, full-time or co-op. If you're working on a hard problem with real stakes, I'd like to hear about it.
