Specializing in Edge Compilation, Parameter-Efficient Fine-Tuning (PEFT/LoRA), Unsupervised Anomaly Localization, and Production MLOps Serving Meshes.
I am an AI Systems & Machine Learning Engineer specializing in parameter-efficient fine-tuning (PEFT/LoRA), edge model compilation (INT8 PTQ/ONNX), grounded retrieval-augmented generation (RAG), and automated SRE telemetry diagnostics.
Currently developing vision pipelines, multimodal diagnostics, and telemetry infrastructure at Agrigate Network Limited. My focus is architecting end-to-end, low-latency microservices with clean API boundaries, rigorous statistical drift detection, and high-performance developer tooling.
Peer-reviewed abstract acceptances and deep learning research contributions.
Investigating deep learning segmentation and reconstruction networks for restoring feature detail and visual quality in ultra-low-light environments.
Generative adversarial translation framework for transferring complex handwriting characteristics across unpaired document datasets without ground-truth alignment.
ONNX Runtime, INT8 PTQ Quantization, Graph Compilation, Latency Profiling
LoRA From Scratch, Rank Adaptation, Cross-Entropy Perplexity, Causal Attention
Canary Traffic Splitting, Shadow Replay, Wasserstein Covariate Drift
PatchCore Anomaly Detection, Sub-pixel Heatmaps, PostGIS Geospatial
Production-grade AI architectures, autonomous developer tools, RAG agents, and MLOps serving meshes.
High-throughput LLM serving engine implementing paged KV-cache memory allocation and continuous iteration batching, delivering an 8.37x throughput gain over static batching.
Autonomous system design review engine auditing architecture anti-patterns (SPOFs, unthrottled queues, unindexed queries) and synthesizing dynamic Mermaid.js flowcharts.
Autonomous SRE incident investigation engine calculating Wasserstein distance anomaly shifts across distributed telemetry, isolating blast radiuses, and outputting runbooks.
Codebase decomposition and cartography engine extracting architectural layers, runtime sequence flows, entry-point manifests, and developer onboarding runbooks.
Multi-speaker conversational diarizer and talk-time analytics engine extracting consensus items, architectural decisions, and synthesizing structured Jira/Linear tickets.
Real-time duplex voice virtual assistant integrating Web Speech API recognition, fast conversational reasoning, and streaming harmonic neural audio synthesis.
Autonomous mock technical interview coach evaluating verbal/text responses across 4 rubric dimensions and synthesizing adaptive probing follow-up questions.
Enterprise customer support RAG agent with hybrid dense/sparse vector retrieval, strict confidence hallucination guards, sub-2ms search telemetry, and verified source citations.
Multi-model registry and dynamic traffic router with SHA256 auditability, probabilistic Canary splitting (0–100%), zero-latency shadow replication, and Wasserstein covariate drift tracking.
Unsupervised industrial defect localization implementing the PatchCore framework, multi-scale convolutional memory banks, nearest-neighbor Euclidean distance scoring, and sub-pixel contour isolation.
Custom Low-Rank Adaptation (LoRA) engine freezing base parameters while training low-rank A/B decomposition matrices on transformer projections, featuring live perplexity convergence tracking.
Model optimization engine compiling PyTorch vision graphs into constant-folded ONNX binaries, executing dynamic INT8 Post-Training Quantization, and profiling p95 latency and MSE tensor drift.
Real-time geospatial danger-zone intelligence and proximity panic alert engine using PostGIS polygon indexing, dynamic risk decay scoring, and asynchronous WebSocket SOS broadcasting.
Mobile safety companion for the CTG Shield engine featuring background GPS telemetry, dynamic polygon containment alerts, one-tap emergency SOS broadcasting, and low-latency WebSocket synchronization.
Zero-network event admission engine using Ed25519 asymmetric cryptography and transactional SQLite replay ledgers to cryptographically prevent ticket forging and replay attacks.
PatchTST-inspired PyTorch engine utilizing sub-series temporal patch tokenization and channel-independent self-attention for multi-step horizon forecasting and anomaly detection.
Autonomous multimodal web navigation agent using visual grounding, DOM bounding box normalization, and asynchronous Playwright perception-action execution loops.
Domain-specific Retrieval-Augmented Generation pipeline built with FAISS vector indexing and SentenceTransformer embeddings for context-grounded advisory generation.
Computer vision microclimate estimator analyzing atmospheric clarity, RGB saturation distributions, and real-time bounding box detections.
Multimodal plant disease diagnostic pipeline fusing deep convolutional leaf feature maps with tabular soil chemistry and ambient environmental telemetry.
Industrial telemetry and growth-curve analytics engine tracking feed conversion ratios (FCR), heat stress coefficients, and ambient ammonia risks.
Available for software & AI engineering roles, research collaborations, and production system consulting.