Founding engineer building production AI systems for messy business workflows: voice agents, LLM workflows, RAG, document intelligence, and data pipelines.
I am the first engineering hire at Tailored AI, where I own architecture and delivery for AI systems that move from ambiguous client workflows to production usage. My work sits at the intersection of applied AI, backend architecture, data pipelines, evaluation, and client delivery.
I work best where the problem is not just "call an LLM", but figuring out the workflow, data model, eval loop, reliability path, and production architecture around it.
- Voice AI systems and explicit STT → LLM → TTS pipelines when reliability matters more than demo latency.
- Agentic workflow systems for logistics, HR/workforce operations, real estate, and customer operations.
- RAG and document intelligence pipelines for freight documents, knowledge workflows, and internal search.
- Production data and analytics pipelines using AWS Lambda, SQS, S3, PostgreSQL, DynamoDB, GCP Vertex AI, and Langfuse.
- Practical ML/data workflows: data cleaning, EDA, feature engineering, XGBoost/Random Forest/ensemble classifiers, demand-pipeline modeling, and evaluation.
- First engineering hire at Tailored AI; helped deliver 6+ AI product deployments across logistics, HR tech, real estate, and call analytics.
- Built freight workflow automation across 80K+ emails, 10K+ loads, and 5 production tenants.
- Built logistics demand automation ingesting WhatsApp and Gmail demand signals, with about 97% average field accuracy in business use.
- Designed and shipped voice-agent architectures, including a production pivot from Realtime speech-to-speech to a controllable STT → LLM → TTS pipeline.
- Won 1st Prize at the All India AI Hackathon by Techolution in 2024 for multimodal RAG on on-device document Q&A.
Most of this work is private/client-facing, so I write about architecture and production lessons rather than exposing implementation details.
Freight AI workflows
Voice and email agents for freight operations, including long-running workflow orchestration, eval-driven prompt improvement, retry logic, and operational analytics.
Voice-first career platform
Designed a multi-service system with Next.js, Django, and Voice agents. Reworked the voice architecture to improve tool-calling reliability, observability, and workflow control.
Logistics demand pipeline and data enrichment
Worked on logistics demand ingestion from WhatsApp/Gmail, structured field extraction, vehicle/location mapping, CRM demand creation, and transporter data enrichment using fuzzy matching and ensemble ML approaches.
Freight document intelligence
Automated freight document classification, splitting, and structured extraction from bills of lading, invoices, and related operational documents using vision LLMs.
Call analytics pipeline
Designed a serverless pipeline for transcription, summarization, sentiment analysis, and intent detection using AWS Lambda, SQS, and GCP Vertex AI.
AI / LLM
OpenAI APIs, Anthropic Claude, Google Vertex AI, RAG, embeddings, prompt engineering, evals, Langfuse, LangGraph, XGBoost, LightGBM, Random Forests, scikit-learn, PyTorch.
Backend & Infrastructure
Python, Django, FastAPI, Node.js, REST APIs, queue-based systems, Docker, Kubernetes.
Cloud
AWS Lambda, SQS, S3, EC2, CloudWatch, GCP Vertex AI.
Databases
PostgreSQL, MongoDB, DynamoDB, Redis.
Frontend
Next.js, React, TypeScript, workflow UIs, dashboards, human-in-the-loop review tools.
Models are only one part of the system. The hard production work is usually:
- turning messy business workflows into clean data and interfaces;
- defining evals before quality becomes subjective;
- logging enough context to debug failures;
- choosing reliability and observability over flashy demos;
- keeping humans in the loop where the cost of an incorrect prediction is high.
I write about these lessons on Tailored AI Substack, including production tradeoffs from deploying voice AI and agentic systems.