All Work

Edge Inference Acceleration · Jan 2025 – Present

• → First peer-reviewed YOLOv11 DPU deployment — accepted as full paper at IEEE ISVLSI 2026.
• → 25.77 FPS on ZCU104 (+49% throughput) and 38.81 ms latency (−33%) vs Intel i5-12650H.
• → Collapsed 9 fragmented DPU subgraphs to 1 via operator substitution (SiLU→ReLU, Softmax→HardSigmoid, matmul→element-wise).
• → Demonstrated hand-curated 250-image calibration outperforms random 1300-image sets under INT8 quantisation.

Algorithmic Fairness Auditing Framework · 2024

• → End-to-end framework auditing ML classifiers across disparate impact, statistical parity difference, equalized odds, and Theil index simultaneously.
• → Comparative mitigation framework enabling evidence-grounded intervention selection across pre-, in-, and post-processing stages.

Knowledge-intensive NLP with sub-second latency · 2024

• → Neo4j GraphRAG retrieval combined with Cache-Augmented Generation (CAG) targeting sub-second inference.
• → KV cache persistence via DynamicCache + pickle serialisation; ablated structured vs. flat RAG baselines.

Privacy-first local AI daemon · 2024

• → Graduated autonomy ladder with HMAC-signed human-in-the-loop approvals.
• → SQLite FTS5 memory distillation; Docker-sandboxed tool execution.

Open-source multi-agent terminal engineering assistant · 2024

• → Gather–Act–Verify loop with specialist subagents: code-explorer, code-architect, code-reviewer.
• → Multi-provider LLM routing (Ollama, Groq, Anthropic, Gemini).

Kaggle "Messy Mashups" — IIT Madras · 2024

• → Ensemble of custom CRNN, fine-tuned ResNet50, and DeiT-Small trained on log-Mel spectrograms.
• → Variable-SNR noise + tempo augmentation; AST via HuggingFace Trainer with W&B experiment tracking.

Hybrid CNN-LSTM with temporal attention · 2023

• → 10-class distraction classification on State Farm dataset (22K+ images) with YOLOv8 backbone + temporal attention LSTM.
• → Mixed-precision training, weighted random sampling, sliding-window temporal context on 4GB GPU.