I am a founder, product architect, and engineer with three decades across ERP, hospitality, logistics, postal systems, warehouse management, and shopfloor control. Today I build the AI layer on top of all of it — voice agents, computer vision, RAG pipelines, and agentic automation — with the discipline that comes from shipping software businesses genuinely cannot afford to have break.
Most AI practitioners have never shipped a system that 500 people depend on daily. I have — for three decades. That's what makes the AI work credible: it integrates with real systems, it survives contact with real data, and it's built by someone who knows what "production" actually means.
Three decades of hands-on software engineering and product leadership across the systems businesses genuinely depend on — ERP, hospitality, logistics, postal, warehouse, and manufacturing operations.
Taking that delivery discipline into LLMs, voice, computer vision, and agentic automation — with a bias toward systems that earn their keep in production, not just in demos.
Each domain below represents systems I've built and maintained in production — not consulting engagements, not advisory work. The AI I build today rests on intimate knowledge of how each of these industries actually operates.
Built a full warehouse management system covering inventory slot control, employee role management (supervisors, pickers, analysts), barcode identification with Code 128 and GS1-128 support, pick-pack-ship workflows, and AI-augmented demand forecasting layered on top of live inventory data.
Designed and built a Kanban-based shopfloor control board for manufacturing SMBs — covering work-order lifecycle management (creation, assignment, progress, completion), multi-site employee scheduling, labor efficiency analytics, productivity reporting, and a computer vision safety monitoring layer using language-grounded NVIDIA models.
Developed product solutions for postal department and private logistics operators — IoT-enabled real-time parcel tracking with two-way communication, route optimization for last-mile delivery, automated delivery workflow management, and AI anomaly detection that flags at-risk shipments before customers need to call.
Three decades with Sage 300 (ACCPAC), Sage CRM, Sage X3, Sage HRMS, and Sage 500 — spanning IC/OE/PO modules, serial and lot tracking, .NET MVC migrations from VB6 and Delphi origins, Java 8-to-11 upgrades, and now AI retrieval copilots and natural language query layers over live ERP data.
Built a full property management system in Delphi — covering front desk, reservations, check-in/out, billing, and F&B integration. Implemented payment tokenization and gateway connectivity (RBSPay, TokenEx) and have since layered a voice-AI receptionist capable of handling phone bookings and inquiries in multiple languages without human intervention.
Designed and deployed a POS system for restaurant operations and built a fully live voice ordering agent — deployed for a real client — capable of handling phone orders in Tamil and English, confirming reservations, routing to kitchen systems, and operating 24/7 without staff involvement. Architected this into a multi-tenant SaaS platform serving any F&B operator.
Built end-to-end tooling for manufacturing operations — BOM and routing management (Sage X3), procurement automation with AI-driven vendor analysis, demand forecasting using Python/Prophet integrated into purchasing workflows, and a shopfloor control board tailored to SMB factories. Also produced full SDLC documentation for maritime software covering accounts, maintenance, and procurement modules. Deep familiarity with the operational realities of Indian MSMEs.
Each entry describes a real system — actual stack, real business problem, actual outcome. No demos, no proofs-of-concept that never shipped.
Built and deployed a fully functional voice ordering agent handling real phone calls — in Tamil and English — for a restaurant client. The agent takes orders, confirms reservations, answers menu queries, and routes to kitchen systems without any human in the loop. Operating 24/7 in production.
Subsequently architected the system into a multi-tenant voice SaaS platform so any F&B or hospitality operator can onboard without per-client engineering. Covered every layer: SIP telephony via Telnyx, TTS/STT latency tuning with ElevenLabs, orchestration on LiveKit, Node.js backend on Railway, system-prompt management per tenant, and a cost model balancing GPU, telephony, and LLM spend. Benchmarked Vapi, Retell AI, Pipecat, and Dograh before selecting the final architecture.
Replaced manual phone order-taking with a 24/7 multilingual agent — zero staffing overhead, no missed calls during peak hours, architectured to scale across operators
Designed and built a full warehouse management system with inventory tracking, employee role management across warehouse functions, site operations, and pick-pack-ship workflows. Implemented barcode identification and validation for Code 128 and GS1-128 formats — the scanning backbone for receiving, putaway, and dispatch.
Layered AI intelligence on top: ML-driven order routing that optimizes picking paths and reduces fulfillment time, dynamic inventory allocation considering customer priority and shipping deadlines, anomaly detection flagging shipments at risk of delay before customers escalate, and computer vision for package condition monitoring at receiving. Also designed predictive safety stock calculations and proactive notification pipelines.
Transforms a conventional warehouse into an intelligent fulfillment operation — predictive signals replace reactive exception management, and mis-pick rates drop measurably
Built a shopfloor control application for manufacturing SMBs: Kanban-style work-order management, multi-site employee and team management, labor efficiency analytics, productivity reporting, and a full activity tracking layer. Deployed on Vercel with bundle optimization for factory-floor devices.
Integrated NVIDIA LocateAnything — a 3B vision-language model — for natural-language object detection on the shopfloor. Instead of training a fixed-class classifier, operators describe what they need to monitor in plain English ("forklift near gate 3", "unattended equipment in zone B"), and the model detects it across camera feeds in real time. Validated on dense industrial scenes with strong spatial reasoning. Also designed a RAG layer for surfacing the right operational knowledge at the right station.
Brings real-time production visibility and AI-assisted safety monitoring to factory floors still running on paper and phone calls — no fixed classifier, no retraining per hazard type
Developed IoT-enabled tracking and workflow systems for a postal department and private logistics clients — real-time parcel tracking with two-way communication between operators and field teams, route optimization for last-mile delivery, and automated workflow management across sorting, dispatch, and delivery legs without manual re-entry between stages.
Applied AI demand forecasting (Python/Prophet) to predict volume spikes and pre-position sorting capacity before peaks arrive. Implemented anomaly detection over live tracking streams that identifies at-risk shipments and triggers proactive customer notifications — replacing reactive exception management with advance warning.
Postal and logistics operations gain live shipment intelligence and AI-generated early warnings — customers are informed before they call, and capacity is positioned before peaks hit
Co-founded a US-registered AI automation venture targeting American SMBs. Designed and built an agentic AR automation pipeline covering the full follow-up and collections workflow: LangGraph for the agent decision graph, FastAPI for the API layer, Celery and Redis for async task orchestration, Supabase for persistence, and a Next.js front-end for the ops team.
Designed the agent's escalation logic, human-in-the-loop checkpoints for edge cases, and the compliance architecture for CCPA and SOC 2 requirements. Evaluated AWS Bedrock, Azure AI Foundry, and OpenRouter for AI hosting trade-offs. Built the cost model and pricing strategy for the venture.
Gives a small SMB finance team the AR follow-up capability of a much larger department — the manual chasing work that quietly drains cash flow is handled by an agent that never forgets a due date
Designed and built a retrieval-augmented generation pipeline that lets ERP users query accounting, inventory, and operational data in plain English — without opening reports or waiting for a specialist. Vector store built on pgvector over live Sage 300 SQL Server data; answers are always source-cited and grounded, never hallucinated figures. Designed to sit beside the system, not replace it.
Any manager or accountant can query the ERP directly — "what's our inventory position on item X?" gets a grounded, cited answer in seconds instead of a ticket to the ERP team
Built a GPT-style language model from the ground up in PyTorch — character-level tokenization, multi-head self-attention, positional encoding, transformer blocks, and training on the TinyStories dataset. A deliberate choice: thirty years of enterprise software meant knowing how to use tools; building from scratch meant understanding what's actually inside them. Also explored small language models (Phi-4-mini via Ollama) for on-premise inference and data-sovereignty use cases.
The reason I can reason about model behavior, attention patterns, and failure modes — not just integrate an API and hope the model behaves the way the docs promise
The biggest gap in industrial AI today is between research-grade vision capability and systems that actually run in warehouses and factories. The work below explores how language-grounded vision changes what's achievable — and affordable — in real operational environments.
Integrated NVIDIA LocateAnything — a 3B vision-language model on HuggingFace — for shopfloor safety monitoring. The model detects any object described in plain English with no fixed class list and no retraining. Tested and validated on real industrial scenes with dense clutter and variable lighting.
Deep operational experience with barcode systems in warehouse and logistics contexts — from identification and encoding validation to integration with WMS workflows at receiving, putaway, and dispatch.
Designed computer vision quality inspection pipelines for manufacturing — combining multiple specialized models for surface defect detection, dimensional analysis, and anomaly classification, with a meta-learner for final decisions.
The core insight across all vision work: language-grounded detection fundamentally changes the economics of industrial vision systems. Operators describe what they need to find; the model finds it — no training pipeline, no ML team required per use case.
The arc that makes the AI work credible and the product instincts trustworthy — thirty years of shipping real systems for real businesses, now applied to the hardest new layer.
Building and shipping AI systems across voice, vision, RAG, and agentic automation. Co-founded a US AI automation venture. Actively exploring SLMs, on-premise inference, and AI integration patterns for businesses where cloud-only is not an option. Leading AI product strategy for an 85-person engineering organization.
Built and led an 85-person engineering and QA organization across six service product lines spanning ERP, hospitality, payments, and CRM. Managed sprint delivery, architecture decisions, client relationships, talent strategy, and the full product roadmap — across India, Singapore, and Washington D.C. operations.
Designed and delivered product solutions for postal department clients and private logistics operators. Built IoT-enabled tracking platforms, route optimization systems, and automated delivery workflow tools — laying the domain foundation that AI anomaly detection and forecasting now build on top of.
Deep architecture and engineering work in warehouse management systems — inventory control, barcode infrastructure, slot management, and pick-pack-ship workflows. Parallel work on shopfloor control, BOM management, and manufacturing operations tooling for SMB clients. The domain expertise that makes the AI work on these systems accurate, not just plausible.
Three decades in ERP and enterprise systems — from VB6 and Delphi origins through .NET MVC, Java, and full-stack web migrations. Built and maintained accounting, inventory, procurement, serial/lot tracking, hospitality PMS, and payment systems for clients across India and internationally. The operating foundation that makes everything else credible.
Range built over thirty years and kept deliberately current. Not a keyword list assembled for a job application — these are tools I've shipped real code with.
That's a genuinely rare combination — three decades of shipping production enterprise systems in ERP, logistics, warehouse, shopfloor, and hospitality, applied directly to building voice AI, computer vision, RAG copilots, and agentic automation. If that's the profile you need, I'd like to talk.