devesh.rawat

New Delhi, India

Devesh Rawat

I build retrieval, verification, and detection systems and the benchmarks that prove they work.

B.Tech Computer Science, Data Science specialization — class of 2027

Portrait of Devesh Rawat

// about

About

I'm a 4th-year Computer Science student specializing in Data Science, currently interested in retrieval, evidence verification, and applied deep learning. At CSIR-HQ I built a system that converts unstructured documents into searchable records and checks claims against retrieved evidence — matching a published SPLADE++ baseline on BEIR SciFact and reaching 98.2% of the leading SciVer reference. Before that, at DRDO SAG, I trained CNNs to classify encrypted image data. Most of my projects start with a working baseline, then go looking for where it silently breaks.

// experience

Experience

Council of Scientific & Industrial Research (CSIR-HQ) May 2026 – Jul 2026

Summer Intern · New Delhi, India

  • Built a system converting unstructured research documents into searchable, verifiable records.
  • Scored 0.711 nDCG@10 on the official BEIR SciFact benchmark — matching the published SPLADE++ baseline (0.710).
  • Reached 98.2% of the leading SciVer reference on claim verification, with 0.832 macro-F1 and 83.7% accuracy across 300 official claims.
Scientific Analysis Group (SAG), DRDO Jun 2025 – Aug 2025

Research Intern · Delhi, India

  • Trained 3 CNN architectures to classify AES-CBC–encrypted CIFAR-10 images, reaching 85% validation accuracy.
  • Built a CustomTkinter GUI to demonstrate the classification pipeline on encrypted data.
  • Collaborated with senior scientists on preliminary findings for defense-technology research.

// projects

Projects

0.61 avg nDCG@10 across four BEIR datasets — 30% above BM25

Bone-Engine — Hybrid Retrieval & Reranking

A hybrid retrieval pipeline combining SPLADE, BGE embeddings, and cross-encoder reranking, validated against published baselines before trusting the results.

  • Reproduced published bge-base-en-v1.5 scores on all four datasets as an independent check before trusting the hybrid result.
  • Measured each stage by ablation: fusion added +1.9 nDCG@10 over dense alone, reranking added +1.6 more.
  • Found the reranker overwriting first-stage ranking costs −5.3 nDCG@10 on claim-style queries — blending at weight 0.3 instead gains +1.6.

Python · SPLADE · BGE · Cross-Encoder

View on GitHub ↗

82.6 mAP@0.5 on PASCAL VOC2007 — ahead of SSD512 and YOLOv3

Neural-Video-Detection-Engine — Anchor-Free Object Detection

An anchor-free object detector trained in 10 GPU-hours on a single 6GB laptop GPU. Found and fixed three silent-correctness bugs the loss curve never showed.

  • CNN backbone, P5-only attention, BiFPN neck, DFL-based decoupled head — the locality-first pattern behind RT-DETR and YOLOv11.
  • Caught a broadcasting fault that inflated box loss N× and a bf16 rounding error breaching a distribution's valid index range, each pinned by a regression test.
  • Moved AES-256-GCM encryption and HMAC-SHA256 audit logging to the checkpoint/API boundary after an in-forward-pass version was found to cost accuracy — now zero mAP cost.

PyTorch · timm · BiFPN · CUDA

View on GitHub ↗

Peak KV-cache memory cut from 4.50 GiB to 0.36 GiB at 32K tokens

Long-Context KV-Cache Compression on Consumer GPUs

Compressed the KV cache during prefill to fit long-context inference on a 6GB GPU, diagnosing a silent accuracy collapse along the way.

  • Found SnapKV retains just 5% of the KV cache while matching full-context accuracy on RULER, across 2,169 runs.
  • Diagnosed a silent failure in query-aware eviction — windows blind to the question collapsed accuracy from 1.00 to 0.00 by chunk 4 — fixed by probing the cache with the prompt's tail first.
  • Traced a 128× memory blowup to a silent kernel fallback, validating a predictive VRAM model against measurement (3.84 vs. 3.86 GiB) to catch it pre-OOM.

PyTorch · Transformers · Qwen3-4B · NF4

View on GitHub ↗

More projects on github.com/deveshr011 ↗

// skills

Skills

Programming & Data
Python · C++ (CUDA) · C · SQL · YAML · Pandas · NumPy · scikit-learn
AI, ML & Retrieval
PyTorch · Hugging Face · Transformers · RAG / SPLADE · BM25 · Cross-Encoders · FAISS
Learning & Evaluation
PPO · DPO · LoRA / QLoRA · CNNs (CSPNet, BiFPN) · BEIR · nDCG@K · MRR
Engineering & Security
Linux · Docker · Kubernetes · Git · CMake · LaTeX · AES-GCM · HMAC-SHA256

// contact

Get in touch

Reach out if you want to talk research, retrieval systems, or anything built here.