Applied AI, LLMs & Intelligent Automation Engineering

DCPL Tech engineered a proprietary, production-tested intelligent invoice reader engine and bulk document conversion pipeline:

• Multi-Vendor Format Adaptation (10+ Separate Vendors): Architected dynamic schema normalization supporting 10 distinct vendor invoice layouts. Each vendor's idiosyncratic table boundaries, font hierarchies, and column alignments are parsed into consistent unified JSON objects.
• Positional & Spatial Pattern Extraction: Curates individual text blocks by calculating spatial positions, relative geometric offsets, and bounding box patterns, reliably associating detached labels with corresponding numerical values even when table borders are absent.
• Multilingual Processing (English & French): Engineered linguistic tokenization models capable of extracting invoice data across languages—accurately recognizing French terms (e.g., Facture, Montant HT, TVA, Total TTC, Date d'échéance) alongside standard English terminology.
• Bulk Batch Ingestion & Serialization: Built asynchronous batch processors that ingest hundreds of PDF and Word files simultaneously, normalize irregular whitespace, and export clean, standardized CSV and JSON records.
• Demonstrated Client Impact: Successfully demonstrated to international enterprise stakeholders (receiving enthusiastic client appreciation), significantly reducing manual bookkeeping overhead and downstream accounting reconciliation delays.

DCPL Tech architected an end-to-end multimodal audio/video processing pipeline integrating cutting-edge computer vision, speech recognition, and deep learning NLP libraries:

• Stream Ingestion & Audio Demuxing via MoviePy: Engineered automated download and stream capture pipelines using requests and moviepy. The software extracts audio tracks from high-definition video files, downsampling and normalizing audio frequencies to optimize acoustic recognition.
• Acoustic Speech Recognition: Employs robust SpeechRecognition engines with ambient noise suppression, speech-segment chunking, and timestamped phonetic-to-text conversion.
• Syntactic Parsing & Entity Extraction via spaCy: Utilizes spaCy linguistic pipelines to identify named entities (attendee names, dates, deliverables, project references) and analyze grammatical dependency trees to isolate commitments.
• Transformer-Based Executive Synthesis: Hugging Face transformers and TensorFlow models condense voluminous meeting transcripts into structured executive summaries: Key Discussion Topics, Decisions Agreed, Action Items with Deadlines, and Follow-up Agenda.
• Online URL & Offline Media Support: Operates seamlessly across local MP4/WAV file uploads or remote video URLs, providing a universal transcription and MoM generator presented directly to enterprise clients.

Our machine learning engineers build end-to-end training and retrieval systems:

• Domain Text Chunking & Tokenization via NLTK: Structured training pipelines tokenize raw domain corpora (260+ enterprise text blocks and vendor documentation) using nltk, formulating optimized contextual chunks and building domain-specific RAG tokens that preserve semantic co-reference.
• Weights & Biases (W&B) Experiment Tracking: Configured automated wandb pipelines monitoring model runs, gradients, and loss metrics in real time. We executed multiple iterative runs tweaking hyperparameters (learning rates, batch sizes, context boundaries), verifying that training loss decreased significantly across 10 complete Epoch cycles.
• Hardware Acceleration with NVIDIA CUDA & cuDNN: Configured low-level CUDA and cuDNN libraries on remote GPU servers, establishing environment paths and optimizing PyTorch GPU cache management to eliminate memory thrashing during long-running training loops.
• RAG Context Engineering & Vector Retrieval: Coupled dense semantic vector embeddings with dense retrieval indices. The model dynamically builds context from prompts and vector databases, providing pinpoint citations and eliminating hallucinations.
• Continuous Dataset Preprocessing: Automated text-to-CSV and text-to-JSON data conversion pipelines continuously transform bulk multi-vendor archives into preprocessed datasets for model ingestion.

DCPL Tech bridges the gap between machine learning engineering and enterprise DevOps:

• Flask Microservice API Bridge: Engineered a lightweight Flask application providing REST endpoints that bridge client web applications with backend ML models. Developed a specialized ragflow context builder function that dynamically extracts vector context from domain corpora and passes the compiled context into the local Ollama execution API for streaming generation.
• Complex Dependency Resolution (340+ Libraries): Managed and resolved intricate version incompatibilities across more than 340 Python libraries in the production requirements manifest—harmonizing CUDA runtime bindings, PyTorch tensor engines, Hugging Face transformers, spaCy models, and web servers into a single hermetic build.
• Hermetic Docker Containerization: Encapsulated the Flask web bridge, ML weights, and execution environments into production-ready Docker containers with cache optimization and rapid container spin-up times.
• Microsoft Azure Cloud CI/CD Pipelines: Integrated directly with client Azure environments (Azure Container Registry and Container Instances), establishing automated CI/CD pipelines that test, build, and deploy new model iterations continuously.
• Agile Engineering & Client Transparency: Delivery operates through disciplined daily scrums, collaborative GitHub pair programming sessions (led by senior engineers Akansha, Debabrata, and Basu), and regular sprint demonstration calls with global leadership (Sheldon, Bruce, Warren), maintaining high stakeholder confidence and clear milestone delivery.

DCPL Tech engineers autonomous browser automation utilities:

• Headless Selenium Orchestration: Developed robust Selenium automation frameworks in Python (PyCharm/VS Code) engineered to resolve complex DOM rendering, explicit dynamic wait states, session cookie management, and multi-step modal interactions.
• Resilient Issue Resolution & Exception Handling: Implemented defensive exception-handling layers (stale element recovery, dynamic locator fallback strategies) co-developed during collaborative technical problem-solving sessions.
• Deep Structured Data Extraction: Harvests unstructured catalog tables, pricing tiers, product descriptions, and downloadable attachments, immediately sanitizing data and feeding downstream analytics engines.
• Micro-Automation Tools: Rapidly builds lightweight task automation scripts that eliminate repetitive internal manual workflows across desktop environments and remote web services.

Our ongoing technical exploration and production implementation spans:

• GGUF Model Quantization: Actively quantizing large language models into modern GGUF formats (4-bit and 8-bit precision). Quantization reduces memory bandwidth constraints and VRAM footprint by up to 70%, allowing high-parameter models to run at blazing token speeds on commodity edge hardware.
• Local Sovereign Inference with Ollama: Deploys and benchmarks local LLM runtimes using Ollama and quantized weights on local developer workstations and client private servers, ensuring zero client data ever leaves the security perimeter.
• Mistral & Hugging Face Model Specialization: Fine-tunes and prompts state-of-the-art open models (Mistral 7B/8x7B, Llama family, specialized encoder-decoders) for domain-specific information extraction tasks with minimal latency and near-zero marginal inference token costs.
• Multimodal Architecture Research: Actively studying and experimenting with Next-GPT-4 multimodal architectures, cross-attention alignment, and unified processing across text, audio, and visual inputs.
• Continuous Research Paper Synthesis: Our dedicated engineering team actively reviews latest arXiv machine learning papers, open-source repositories, and parameter-efficient fine-tuning (PEFT/LoRA) techniques to rapidly incorporate production-ready breakthroughs into client deliverables.