LinkedIn · Email

Question. Can a hallucination detector still read an LLM's hidden states after the model is quantized?

Approach. Linear truthfulness probes on Llama-3.2-1B and Qwen2.5-1.5B at FP16, INT8 and INT4. 6,049 true/false statements, topic-held-out splits, byte-identical inputs across precisions.

Result. Probe accuracy doesn't change under quantization (p = 0.38), and moving a probe across precisions costs under 0.02 AUROC.

Python PyTorch Transformers bitsandbytes scikit-learn

Question. Can a model classify leukemia subtypes from a blood smear and show why?

Approach. A CNN ensemble for the whole image plus a GNN over segmented cells, fused with cross-modal attention. Grad-CAM and GNN attention show where it looked; MC dropout gives confidence intervals.

My part. The data pipeline: four public datasets, class imbalance, train/val/test manifests, and preprocessing (stain normalisation, cell segmentation) that feeds both pathways. I also presented the idea.

PyTorch PyTorch Geometric OpenCV FastAPI · Team project at UPES with Brajraj Singh Pathania, Piyush Bharadwaj and Ayushmaan Singh, guided by Prof. Gouranga Duari

Question. How do you send thousands of scheduled emails exactly once, when the server can die mid-campaign?

Approach. Keep the schedule in Redis and Postgres, not in a process. BullMQ delayed jobs, per-sender hourly limits enforced by an atomic Lua script, and jobs over the limit move to the next hour instead of failing.

Result. 1,000 emails scheduled in 1.8 s, spaced exactly 2.0 s apart, and nothing lost or duplicated across restarts.

TypeScript Express PostgreSQL Redis BullMQ Elasticsearch Next.js

Question. How cheaply can you turn a rough instruction into a good prompt, and what will that prompt cost to run?

Approach. Two stages on different models: a small model finds the gaps in the draft, a frontier model writes the fix. Response caching, prompt caching, and a free cost estimate before anything is spent.

TypeScript Next.js Claude API

Question. How risky are crypto markets, which phase of the cycle are they in, and do they move with traditional assets?

Approach. Volatility and VaR, GARCH, market-cycle phases, rolling correlation against equities, gold, the dollar and oil, and walk-forward ARIMA checked against a random-walk baseline.

Python pandas statsmodels Plotly Streamlit

Question. Can a resume be scored against a job description the way a careful reader would?

Approach. Text extraction, then Sentence-BERT similarity plus skill overlap, ranked into a match score.

Python Sentence-BERT Streamlit