IceCache: Semantic Paging for 1M-Token Context

A KV-cache management layer that clusters semantically related tokens and integrates with PagedAttention, enabling long-context inference on consumer GPUs without the accuracy cliff that naive offloading suffers.

Technical Approach

Implemented as a new BlockManager subclass inside vLLM, IceCache computes cluster assignments over cached keys using mini-batch k-means on hidden states, defines cluster-granular offload/retain policy (keeping centroids on GPU, offloading members to CPU), and rehydrates clusters on-demand based on the current query's nearest-centroid match via FAISS.

Key Results

• 128K-context inference on a single 24GB RTX 4090 with <3% accuracy loss on LongBench
• Custom Triton kernels for online k-means centroid updates achieving 1.7× throughput vs. baseline CPU-offload on chain-of-thought reasoning workloads
• Open-sourced as a vLLM fork with Terraform one-click AWS deployment

Token-Budget-Aware Pool Router

A traffic router in front of heterogeneous vLLM instances that estimates each incoming request's token budget and dispatches to a pool sized for that budget — the missing Kubernetes ingress for LLM serving.

Technical Approach

Self-calibrating per-category bytes-per-token estimator that classifies requests into token-budget buckets using request features + a running BPT ratio, then routes to differentially-configured vLLM pools. Built as a Go/Rust gateway with KEDA for pool autoscaling, Prometheus for routing telemetry, and Locust for load testing with realistic request mix traces.

Key Results

• Eliminated OOM preemptions and increased effective concurrency by 4.2× on production traffic replay
• Self-calibrating BPT estimator converges within 500 requests per category
• Reduced GPU cost per million tokens by 61% while keeping p99 TTFT under 400ms

STARS: Runtime Skill Firewall for Agents

A low-latency safety layer that scores every tool/skill invocation against the user request + runtime context at call time, blocking or escalating risky invocations before they execute. Continuous risk estimation for the MCP era.

Technical Approach

Request-conditioned continuous risk scorer: a distilled 600M classifier ingesting (user_request, candidate_skill_signature, runtime_context_features) → [0,1] risk. Fine-tuned on synthetic risky/benign invocation traces, deployed via ONNX Runtime for sub-10ms inline scoring. Built as MCP middleware with Redis caching and Kafka audit stream.

Key Results

• p99 <8ms scoring latency, blocking 91% of policy-violating invocations
• Continuous retraining from human-labeled false positives improved precision from 82% → 96% over 2 weeks
• Integrated with MCP-compliant agents (Claude, Qwen-Agent) as transparent proxy, adding <2% overhead on end-to-end task completion time

World Model Lens

An active research project at Elemental Research Lab investigating how language models build internal world representations — and what tools we need to inspect them.

Research Questions

• What features do language models use to represent world state?
• How do these representations compose across layers?
• Can we build reliable probes for world-model quality, and use them for safety evaluations?

This work sits at the intersection of mechanistic interpretability and AI safety — understanding the internal structure of models is a prerequisite for building systems we can trust.

Pyramid Hierarchical Spatial-Spectral Transformer

A novel Transformer architecture for hyperspectral image classification that achieves state-of-the-art results while being significantly more parameter-efficient than existing methods.

Key Results

• 99.21% accuracy on Kennedy Space Center dataset, outperforming 10+ baselines by 1.8–6.3 percentage points
• Ranked top 3% on HSI classification leaderboard
• Reduced model size by 30% (2.1M → 1.47M parameters) while maintaining accuracy
• Enabled edge deployment, saving $12K annually in computational costs
• 12+ citations validating research impact

Graph-Based Sentence Selection with Transformer Fusion for Text Summarization

Published in the Journal of Computer Engineering and Technology (JCET), Vol. 7(1), 2024 — cited 15 times. A hybrid summarization framework that pairs graph-based extractive sentence selection with Transformer-based abstractive fusion.

Approach

Builds a sentence-similarity graph using contextual embeddings, ranks sentences via a modified TextRank, and feeds the selected subset into a Transformer encoder-decoder for fluent rewriting. The two stages are jointly tuned so the extractor learns to produce inputs the abstractor can compress most effectively.

Key Results

• Synergistic gains over either standalone graph or standalone Transformer baselines on CNN/DailyMail and XSum
• Reduced redundancy in long-document summarization while preserving factual coverage
• Demonstrates that graph-based selection is a useful inductive bias for Transformer summarizers

Sustainable NLP: Parameter Efficiency for Resource-Constrained Environments

Published in the Journal of Computer Engineering and Technology (JCET), Vol. 7(1), 2024 — cited 11 times. A systematic study of parameter-efficient fine-tuning and compression methods for deploying NLP models in low-compute settings.

Approach

Comparative evaluation across LoRA, adapters, prefix tuning, knowledge distillation, and post-training quantization on a suite of representative NLP tasks. The study isolates the trade-off between trainable-parameter budget, inference cost, and task performance, and proposes practical decision criteria for choosing a method given a deployment constraint.

Key Results

• Showed parameter-efficient methods can match full fine-tuning within 1–2 points on most benchmarks while training <1% of parameters
• Quantization-aware adapters offered the best accuracy/footprint balance for edge deployment
• Established practical guidance for sustainable NLP under tight compute budgets

High-Performance PDF Retrieval with RAG

An end-to-end retrieval-augmented generation pipeline optimized for large-scale PDF document retrieval.

Key Results

• 40% latency reduction (2.1s → 1.3s) via intelligent chunking, Sentence Transformers, and FAISS indexing on 10,000+ PDFs
• Sub-second query response (mean: 0.8s) with 99.1% availability sustaining 500+ concurrent queries
• Reduced hallucination by 35%, improving downstream LLM accuracy from 82% → 88%

Efficient CAPTCHA Image Recognition Using CNNs and LSTM

Published in the International Journal of Scientific Research in Engineering and Management, 8(3) — cited 28 times. A hybrid CNN-LSTM architecture for solving variable-length character CAPTCHAs.

Approach

The CNN extracts local visual features from the input CAPTCHA image, the LSTM models sequential character dependencies along the recognized string, and the two networks are jointly trained end-to-end. This pairing handles distorted, noisy, and overlapping characters more robustly than CNN-only or template-matching baselines.

Key Results

• Outperforms CNN-only baselines on standard CAPTCHA benchmarks under noise and distortion
• Handles variable-length character sequences without needing per-character segmentation
• One of the most-cited works from the lab on visual sequence recognition

End-to-End Neural Embedding Pipeline for Large-Scale PDF Retrieval

Published in the International Journal of Advanced Research in Computer Science — cited 13 times. A retrieval pipeline for large PDF corpora that emphasizes both quality and scalability.

Approach

PDFs are parsed and chunked with a strategy designed to preserve semantic coherence at chunk boundaries. Each chunk is embedded with a Sentence Transformer model, and the resulting vectors are indexed using a distributed FAISS configuration that supports horizontal scaling across shards. At query time, semantic search returns the most relevant chunks for downstream summarization or generation.

Key Results

• Sub-second nearest-neighbor retrieval at corpus scale
• Robust to messy real-world PDFs (mixed layouts, scanned pages, multi-column text)
• Foundation for downstream RAG applications over PDF collections

Advanced Underwater Image Quality Enhancement

arXiv preprint arXiv:2410.14285, 2024 — cited 9 times. A two-stage pipeline that combines super-resolution with multi-scale Retinex defogging to recover quality from degraded underwater imagery.

Approach

Stage one uses a hybrid Super-Resolution CNN to restore detail and resolution lost to scattering and absorption. Stage two applies multi-scale Retinex normalization to correct underwater color casts and remove residual haze. The combined pipeline produces visually clearer images suitable for downstream marine vision tasks like object detection and habitat monitoring.

Key Results

• Significant qualitative and quantitative improvement on standard underwater image enhancement benchmarks
• Better color fidelity and contrast than either stage applied independently
• Suitable as a preprocessing front-end for marine object detection systems

Systematic Exploration of Dialogue Summarization Approaches

arXiv preprint arXiv:2410.15962, 2024 — cited 2 times. A reproducibility-focused study of dialogue summarization methods on standard benchmarks.

Approach

Reproduces extractive, abstractive, and hybrid dialogue summarization approaches under a consistent evaluation protocol on SAMSum and DialogSum. Isolates where each family of methods performs best (informal chat vs. structured meetings, short vs. long), and proposes methodological refinements for fair comparison — including consistent preprocessing, evaluation metrics, and ablation procedures.

Key Results

• Clarifies which dialogue summarization methods generalize across benchmark types
• Surfaces reproducibility gaps in prior reported results
• Proposes a standardized comparison framework for future work

Never Lost in the Middle Again: Teaching LLMs to Care About the Center of Long Documents

Available at SSRN 5877962, 2025 — preprint. Empirically diagnoses and mitigates the "lost-in-the-middle" failure mode, where LLMs underweight information located in the center of long documents.

Approach

Quantifies position bias across context windows on long-document QA and summarization, then introduces training-time interventions that explicitly upweight middle-context relevance during fine-tuning. The approach is model-agnostic and integrates with standard instruction-tuning pipelines.

Key Results

• Recovers performance on long-document QA benchmarks where baseline models systematically miss middle-context evidence
• Improves summarization faithfulness for long inputs
• Maintains overall performance on short-context tasks — no negative transfer