Asheesh Bhardwaj

Designed and deployed an Agentic AI-based intelligent document processing system for automated loan approval platform, processing over 2M+ construction loan documents/month. Achieved a 45% reduction in manual review time through a comprehensive MDLC implementation leveraging hybrid OCR + LLM + ML model architecture with multimodal validation and intelligent workflow automation. Implemented robust data collection pipelines aggregating multi-format document sources (PDFs, images, invoices, certificates) with automated data quality validation and preprocessing workflows. Engineered feature extraction processes from unstructured documents using hybrid OCR-LLM pipeline, handling missing values, noise reduction, and format standardization across 2M+ monthly documents. Developed hybrid neural network architecture combining CNN layers for visual document analysis with transformer-based LLMs for semantic understanding. Implemented backpropagation optimization using Adam optimizer with ReLU activation functions for hidden layers and sigmoid for binary classification outputs. Applied advanced feature engineering techniques including document embedding generation, semantic similarity scoring, and multimodal feature fusion. Established comprehensive model evaluation framework using cross-validation, precision-recall metrics, and confidence scoring thresholds. Achieved >90% alignment accuracy through iterative model tuning, hyperparameter optimization, and ensemble methods. Implemented A/B testing protocols to validate model performance against baseline OCR systems. Built production-ready ML pipeline using FastAPI for real-time inference with distributed agent controller architecture. Implemented MLOps best practices including model versioning, automated retraining pipelines, and performance monitoring dashboards. Deployed scalable system handling 2M+ documents/month with <2s latency and 99.9% uptime. Auto-generated missing line-item descriptions using multimodal analysis of inspection photos and architectural plans, achieving >90% semantic accuracy through feature-engineered document embeddings. Developed contextual recommendation engine using NLP-based user profiling and policy matching algorithms, reducing approval delays through intelligent SOP automation. Implemented semantic validation layers ensuring coherence between budget data and supporting documents via transformer-based similarity scoring. Built semantic similarity matching engine with custom-trained embeddings, enabling multi-vendor document format support through advanced feature engineering and fuzzy matching algorithms with confidence-based threshold tuning. Developed content recommendation engine clustering inspection photos using CNN-based visual feature extraction and cosine similarity scoring for contextually relevant photo-to-line-item matching. Implemented ensemble models combining structured and unstructured data matching with anomaly detection algorithms for overbudget prediction and claim verification. Engineered fallback heuristics with probabilistic confidence scoring, ensuring robust performance across edge cases through systematic model evaluation. Built audit trail system with explainable AI features providing granular ML/LLM decision transparency for compliance reviews and model interpretability. Implemented real-time feedback loops with active learning capabilities, enabling continuous model improvement through human-in-the-loop validation and automated retraining pipelines.

Optimized data backup and recovery architecture for a SaaS platform handling 10M+ API requests/day; reduced latency by 35% using Amazon S3 Multipart API with dynamic chunking, deduplication, and parallel uploads, improving throughput by 40%. Ingested API logs, system metrics, and incident reports from SharePoint, OneDrive, and Teams via AWS SQS and Lambda. Ensured data completeness and quality with schema validation and automated error-handling pipelines. Cleaned and normalized unstructured logs and metrics. Engineered features such as error frequencies, response-time distributions, and dependency-graph embeddings to feed ML Model development. Handled missing data and outliers with imputation and anomaly-aware filters. Built unsupervised ML pipelines (DBSCAN, LSTM) for pattern detection, using backpropagation with gradient descent to update weights. Fine-tuned transformer LLMs on historical incident summaries and logs for resolution suggestion. Conducted hyperparameter tuning (learning rate, batch size) and optimization with Adam optimizer, minimizing the Loss Function that measures prediction error. Assessed anomaly detectors via precision, recall, and ROC-AUC metrics. Validated LLM recommendations with human-in-the-loop A/B testing and confidence-score thresholds. Implemented Guardrails for LLMs & traditional ML models, including output sanitization, rate limits, and fail-safe fallbacks. Deployed real-time inference endpoints via FastAPI and AWS Step Functions, achieving <200 ms response and 99.9% uptime. Integrated MLOps pipelines for continuous monitoring of drift, data-quality alerts, and automated retraining. Enabled self-healing workflows: feedback loops capture analyst overrides, trigger auto-retraining, and refine labels via weak supervision and active learning. Detected silent failures and regressions early, reducing incident spikes by 50%. Resolution Recommendation Engine: Top-N remediation suggestions via Qdrant vector search; auto-resolves 1K+ incidents/month, cutting MTTR by 60%. NLP-driven summaries translate cryptic logs into human-readable insights for L1/L2 teams. Preemptively flags high-risk backup tasks; continuous learning improves accuracy by 20% over six months.

Developed an IoT solution for UWC Oil Wells to monitor 100K+ devices, improving uptime by 30%. Processed 500GB/day telemetry and enabled predictive maintenance with real-time analytics. Designed and implemented reusable software components, utilized across 3 projects, reducing development effort by 40% and accelerating delivery timelines. Built a robust data pipeline to collect, process, and analyze telemetry data from diverse endpoint devices, ensuring high scalability and fault tolerance. Enabled seamless multi-protocol connectivity (e.g., MQTT, HTTP) for various IoT devices, ensuring compatibility and smooth integration into the existing system. Delivered a comprehensive solution that improved data visibility, reduced downtime, and supported proactive issue resolution.

Designed and developed engaging Casino Slot Games for Web, Facebook, and Mobile platforms, driving user retention and cross-platform compatibility. Successfully implemented Super Mode feature, enhancing user engagement and delivering an immersive gaming experience. Developed an encryption and security module leveraging industry-standard encryption algorithms to safeguard sensitive game data and prevent unauthorized access. Optimized game performance for low-end mobile devices by implementing efficient rendering techniques, memory management, and asset compression, improving load times and gameplay smoothness. Applied advanced game development techniques, including state management, event-driven architecture, and real-time animations, ensuring seamless gameplay.

Architected and implemented a high-performance Data Intelligence Portal to aggregate and analyze critical device information in Naval warships, ensuring real-time access to mission-critical data with ultra-low latency. Reverse-engineered complex communication systems to decode proprietary protocols and optimize data transmission, improving system reliability and operational efficiency. Spearheaded a modernization initiative to upgrade legacy technology infrastructure, replacing outdated systems with cutting-edge solutions at just 2% of the total cost, achieving significant cost savings while enhancing functionality and scalability. Developed high-frequency, low-latency C++ modules for seamless communication and processing of real-time data streams, ensuring rapid response times critical for Naval operations. Designed and implemented an immersive graphical interface for intuitive visualization of critical ship data, providing indispensable situational awareness for warships. Leveraged advanced C++ techniques, including multi-threading, lock-free data structures, and real-time event processing, to optimize system performance in resource-constrained environments. Collaborated with cross-functional teams to integrate the system with onboard sensors, radars, and communication devices, enabling robust and fault-tolerant data handling.

Designed and developed a Warm Standby System to ensure seamless business continuity by maintaining a fully operational secondary environment capable of rapid activation in the event of primary system failure. Implemented high-performance C++ modules to synchronize data and processes between primary and standby environments with minimal latency, ensuring near-real-time availability. Utilized distributed systems principles and fault-tolerant architectures to enhance system reliability and scalability. Integrated the system with high-throughput messaging queues and low-latency data replication techniques, achieving consistent data integrity across environments.

Designed and developed a real-time data integration and visualization system for the Indian Navy, leveraging modern C++11 features to enhance code performance, reliability, and maintainability. Seamlessly integrated and displayed real-time data from various critical sensors, including radar, sonar, and weapon systems, on a centralized platform, enabling comprehensive situational awareness for naval commanders. Utilized C++11 features such as smart pointers, lambda functions, and multithreading to implement efficient data processing pipelines for handling high-frequency sensor inputs with minimal latency. Engineered a modular and extensible architecture, leveraging RAII principles and template programming to ensure optimal resource management and scalability.