Models @ HuggingFace https://huggingface.co/aifinitydigital

Github https://github.com/aifinityresearch

LinkedIn https://www.linkedin.com/in/dharanibabu2025

Cloud Contributions https://medium.com/@dharanibabu_68045/a-free-video-conferencing-solution-on-amazon-cloud-b90f607326db

• Led end-to-end AI architecture and solutioning for a Product Safety Incident Classification (PSIC) platform to automatically assess severity of customer complaints across Smoke Detectors, CO Detectors, and Fire Extinguishers, reducing manual legal review effort.

  • Designed a batch inference AI pipeline to process large volumes of residential customer complaints and classify incidents into severity tiers (standard, escalation-risk, potential legal exposure) using LLM-based reasoning.

  • Architected an LLM-driven classification system leveraging hidden Chain-of-Thought (CoT) to infer latent risk signals without exposing reasoning artifacts, aligning with legal and compliance requirements.

  • Evaluated multiple pre-trained transformer and foundation models — including Gemma 3, Mistral, Llama 3.2, and gpt-oss:20b — across accuracy, reasoning depth, latency, and GPU memory footprint, selecting gpt-oss:20b for production deployment.

  • Implemented a private, on-prem-style LLM deployment using Ollama within an AWS private subnet, ensuring data residency, IP protection, and zero external data leakage.

  • Deployed the production inference stack on AWS g5.xlarge GPU instances using a Deep Learning AMI, optimizing for throughput, model load time, and batch processing efficiency.

  • Enabled early detection of legal and safety escalation risks, significantly reducing manual investigation time and improving proactive compliance posture for product safety incidents.

• STELLAW | AI-Powered Legal Research for Indian Jurisprudence

• Directed the end-to-end AI architecture of STELLAW, an AI-powered legal consultation and research platform purpose-built for Indian jurisprudence, serving professional lawyers and legal researchers.

  • Designed and implemented a Retriever-Augmented Generation (RAG) pipeline by embedding frequently referenced Indian statutes, case laws, and legal doctrines into vector embeddings to enable high-precision semantic retrieval.

  • Architected a multi-model inference strategy, combining Phi-3 Mini for low-latency tasks with DeepSeek-R1 for complex legal reasoning and long-form judgment analysis.

  • Built a scalable FastAPI-based AI backend, integrated with Groq for accelerated inference on non-sensitive workflows.

  • Implemented agentic AI workflows using LangGraph to support advanced features such as judgment analysis, pleading assistance, and multi-step legal reasoning.

  • Established data privacy–aware architecture, routing sensitive legal analysis to privately hosted models while allowing non-sensitive queries to leverage managed inference.

• CapitalMind | LLM-Powered Capital Markets Research Engine

  • Architected CapitalMind, an AI-native investment intelligence platform enabling investment bankers, equity analysts, and capital markets teams to analyze institutional behavior using natural-language queries.

  • Designed a domain-specific LLM fine-tuning pipeline, training a DeepSeek-R1:14B model using 5+ years of SEC Form 13F filings, IPO allocation data, and institution-mapped financial datasets.

  • Implemented LLM orchestration and agent-based workflows using LangChain and LangGraph to support multi-step analytical reasoning.

  • Enabled advanced analytics such as fund flow tracking, quarterly position shifts, and peer benchmarking across major institutional investors (e.g., BlackRock, Citadel, Vanguard).

  • Deployed a privately hosted LLM inference stack using Ollama on AWS private cloud infrastructure, leveraging g4dn.xlarge GPU instances with NVIDIA PyTorch–enabled AMIs to ensure data confidentiality and predictable inference latency.

  • Balanced model size, cost, latency, and accuracy, opting for a fine-tuned small model over larger foundation models to optimize institutional-scale usage economics.

InHouse CoPilot Solution

• Architected and deployed an in-house AI Copilot for enterprise .NET / C# codebases, enabling secure, context-aware developer assistance without reliance on external SaaS tools.

• Integrated Cline (RooCode) with an internal MCP Server, enabling structured and controlled context exchange between IDE tooling and backend LLM services.

• Deployed a privately hosted LLaMA 3.2 70B model for deep cross-file code understanding and architectural reasoning, ensuring full source-code confidentiality and IP protection.

• Implemented the LLM serving layer using Ollama on AWS EC2 private cloud infrastructure, optimizing inference through quantization and GPU-aware deployment strategies.

• Designed the solution to operate entirely within a private network boundary, with no external API calls, aligning with enterprise security and compliance requirements.

• Enabled internal engineering teams to perform contextual code search, refactoring suggestions, and architectural Q&A over large C# codebases, improving developer productivity while maintaining organizational control over models and data.

TASCUA - Tasconnect Universal Adapter - IPaaS Platform

• Led the design and implementation of an in-house Integration Platform as a Service (iPaaS) built on WSO2 (Apache 2.0 licensed), enabling scalable and low-code enterprise system integrations.

• Architected a low-code integration framework to connect heterogeneous ERP and SaaS systems including SAP ECC, SAP S/4HANA, Oracle NetSuite, and Salesforce, reducing custom integration development effort.

• Implemented high-volume invoice processing pipelines, successfully processing 100K+ invoices from SAP systems using WSO2 Micro Integrator, ensuring reliability and transactional integrity.

• Hardened the open-source integration stack for production readiness by identifying and remediating Apache-level security vulnerabilities, aligning the platform with enterprise security standards.

• Designed and enabled containerized deployments of WSO2 Micro Integrator, deploying the integration runtime on AWS ECS Fargate, tailored to client-specific scalability and isolation requirements.

• Introduced observability and distributed tracing across integration services by designing and implementing AWS X-Ray using a sidecar pattern, significantly improving visibility into integration flows, latency bottlenecks, and failure points.

• Established operational best practices for monitoring, scalability, and fault isolation across containerized integration workloads in AWS.

ML Engineer | Systems Analytics Lead - Capacity Planning & Forecasting 

Large Retail Banking Platform | On-Prem Data Centers

• Led the design of an ML-driven Capacity Planning Engine (CAPE) to determine optimal infrastructure capacity for ~8,500 Linux servers supporting a high-traffic retail banking mobile application with ~6,000 concurrent user sessions per second.

• Formulated the problem as a statistical modeling and what-if simulation exercise, enabling infrastructure teams to assess whether existing CPU and memory resources could sustain 2×–3× traffic surges during peak events such as seasonal and holiday sales.

• Aggregated and aligned hourly peak user traffic metrics with corresponding CPU and memory utilization across a one-year historical window, creating a clean analytical dataset for capacity modeling.

• Modeled the relationship between user traffic (independent variable) and CPU/memory utilization (dependent variables) using Ordinary Least Squares (OLS) linear regression, producing an interpretable relationship of the form y = mx + c for capacity extrapolation.

• Evaluated linear regression implementations using scikit-learn and statsmodels, selecting statsmodels OLS for its statistical interpretability, confidence intervals, and diagnostic outputs.

• Built time-series forecasts for user traffic and infrastructure utilization to enable forward-looking what-if simulations under peak and stress conditions.

• Prototyped equivalent traffic-to-resource utilization models in PyTorch, experimenting with gradient-based linear and non-linear formulations, and compared results against OLS to validate model simplicity versus predictive stability.

• Validated model reliability using p-values, R², residual analysis, and train/test splits, ensuring statistical soundness before using predictions to inform infrastructure scaling and procurement decisions.

• Technical Manager for 20+ mil SG Mobile Banking replatform from KONY to native iOS/Android mobile banking application designed with rich UI/UX experience
• Implementation  3DS ( 3D secure ) , Faster Payment System  , eStatement enhancements , IAM product integration
• Ownership of TDD , BDD , DevOPS and Test Automation ( coverage 78%)

• Responsible for Capacity Planning for new LifeStyle2.0 CARDS+ CAT1 Mobile Application Platform 
• Responsible for Application and Infra Deployment Readiness

• Forex Time Deposit , 1FA for low and medium risk transactions
• Solution Architect for Implementation of "6- tier" and "3-tier " Zero Down Time resiliency projects for HK & SG Mobile Banking

• Project manager for Cards Remodeling project 
• Responsible for deliverables on developments on ETL , SORs , Integration 

• Responsible for reverse engineering approach to re-engineer front end applications of Branch IT platform ;De-compilers , porting legacy VC++ 2.0 to .NET framework 
• IBM MQ Client Code migration to .NET framework

• Maintenance /enhancement of Repo Trading Platform- GTRepo - Fixed
  Income Prime Brokerage division 
• 24x7 L2-L3 support for  settlement systems flow of high worth , sensitive trades 
• End to end orchestration for Decommission 

• Solution Architect cum Technical Manager  for  implementing Remote IP . Remote IP helps removing requirement of Public IPs in the customer premises , which was erstwhile required by all Alcatel Business partners. DTMF based SSH tunneling solution for service connectivity to remote PBX systems
• Responsible for L3 maintenance of OXO product line, solutioning of new enhancements - MyIC mobile development 
• Full EMEA ownership with L2 & business partners 

• Tech lead for PIMPhony(softphone) , Omni PCX Management Console , TAPI and Open Telephony Server , Static and dynamic memory
  management initiatives
• Ensuring adherence to SLAs on maintenance defects

• Enhancement  of OmniPCX Management Console ,Open
  Telephony Server and TAPI for soft phone interface with PBX.
• Development for backward compatibility aspects using gSoap services
  and conversion of FTP / wininet into HTTP/ WinHTTP protocol
• Usage of DevPartner Visual Studio for memory leaks detection 

• Developer for OmniPCX Management Console - on Windows MFC , Win32 , FTP & TCP 
• Won "FIRST CENTURY" award for having fixed first 100 defects  

• Design and development of Remote Monitoring WMI Server , developed
  using TCP/Win32 API 
• Custom Multithreaded Threadpool - custom Message Queue  - Pollers