AI / ML engineer, Hyderabad
Most LLM projects stall in the same place. Retrieval that looked fine on a sample folder falls apart on ten years of real documents. Agents loop, or quietly fail without anyone noticing. Latency and token spend go somewhere nobody budgeted for.
That gap is what I do. Thirteen years building and running software, three of them as CTO, now spent entirely on getting language-model systems past the prototype and keeping them alive once they are there.
Four things, and I turn down work outside them. Depth in a narrow band is the reason to hire an independent rather than an agency.
RAG systems that hold their accuracy when you point them at the whole archive instead of the sample. Chunking and parsing for messy source documents, hybrid and graph retrieval, reranking, grounding checks, and an evaluation harness so you can tell whether a change made things better or worse.
Google Agent Development Kit (ADK) LangChain / LangGraph Haystack Qdrant Graph RAG Bedrock
Multi-step agentic systems built to fail safely: bounded loops, tool contracts that hold, retries and fallbacks, human handoff where it belongs, and traces detailed enough to reconstruct any run after the fact. Built with MCP and A2A where they earn their place.
LangGraph MCP / A2A Claude & OpenAI agent SDKs tool design
Conversational pipelines that run under a second end to end, which is a different engineering problem from a chat box. Streaming, barge-in, model and provider routing, latency budgeting across every hop, and graceful behaviour when one hop is slow.
vLLM LiteLLM streaming pipelines telephony integration
Two weeks, fixed fee, no commitment beyond it. I read the code and the traces, talk to the team, and hand back a written assessment: what is actually wrong, what it costs to fix, what to stop doing, and whether the thing is worth continuing at all. Some clients stop there. That is a fine outcome.
Often the fastest way to find out if we should work together
Fraud detection system
For bank transactions to detect fraud in real time. The system uses a combination of machine learning models and rule-based algorithms to flag suspicious transactions, and it integrates with the bank's existing infrastructure for seamless operation.
OCR
Optical Character Recognition (OCR) system to extract text from scanned documents and images. The system is designed to handle various fonts and layouts, providing accurate text extraction for further processing and analysis.
Chatbot for finance system
AI-powered chatbot for financial institutions to assist customers with account-related queries, providing instant support and reducing the burden on human agents.
Quant finance
A multi-agent system for quantitative finance to analyze stock market data and predict buy, hold, or sell signals.
Live multilingual translation
Ekip IT Solutions
2025–2026
Architected an LLM-powered translation system for live conversation, where anything above a second of delay makes the product unusable. The work was mostly latency budgeting: choosing what could be streamed, what could be predicted ahead, and what had to be cut from the path entirely.
Sub-second end-to-end latency
Telephony agents in production
Ekip IT Solutions
2025–2026
Designed and shipped voice agents handling real customer calls with natural language understanding and context carried across the conversation. Alongside them, agentic systems that run multi-step workflows on their own, make decisions mid-call, and adapt when the conversation goes somewhere the script did not anticipate.
Live customer traffic
CTO, two SaaS products
BAWES · Kuwait, remote
2016–2025
Nine years with one group, the last three as CTO, owning engineering for Plugn (eCommerce SaaS) and Studenthub (student job portal). Set technology strategy, hired and ran the team, moved the platform onto microservices so AI could be introduced without rewriting everything, and kept a public SaaS standing through DDoS, spam and brute-force attacks.
10× web application performance
Achieved through React/Vite upgrades, vendor chunking, selective module loading, service workers, Cloudinary image handling, Algolia search and CloudFront edge caching — with AWS auto-scaling and Go for CPU-bound work on the backend.
Products shipped
Selected
Scope and price agreed before anything starts. No open-ended hourly billing.
2 weeks · fixed fee
A written assessment of an existing AI system: failure modes, cost drivers, what to fix first, and an honest read on whether to continue. Ends cleanly with no obligation to go further.
6–12 weeks · per project
I build the system and hand it over working, documented, instrumented and with evals your team can run without me. Handover is part of the scope, not an afterthought.
2–3 days/week · monthly
Ongoing technical ownership for teams building AI without a senior person who has shipped it before. Architecture, code review, hiring input, and saying no to the expensive ideas.
I have worked remotely for companies outside India since 2016, including three years running engineering for a Kuwait-based group. The practical friction of cross-border work is already solved on my side.
Google Cloud’s Generative AI Leader certification covers the part clients usually care about most: deciding where generative AI belongs in a business, what it will cost to run, and where the governance and risk sit. It pairs with the engineering rather than repeating it.
Verify this credentialEvery badge below links to the issuer for verification.
Google Cloud
Verify on Credly
Amazon Web Services
Verify on Credly
Microsoft
Verify on Microsoft Learn
Also completed: Hugging Face — Fine-tuning Language Models, LLM Fundamentals, and Fundamentals of MCP. Bachelor of Engineering in Computer Engineering, Government Engineering College, Rajkot.
Thirty minutes, no pitch. Describe the system and where it is stuck, and I will tell you what I think is happening and whether I am the right person for it. If I am not, I will say so.
