Headline

AI Engineer | Synthetic Data Gen @ SkanAI | Scaling AI Systems (LLMs Inference, RAG, Inference, Observability) | 1,600+ Voice sessions · 7x Throughput · $1.3M Savings

About

I build AI systems that don’t break in production. Most AI work looks good in demos and falls apart under real load—latency spikes, costs explode, and failures are impossible to debug. My work sits on the other side of that line: making LLM and voice systems **fast, reliable, and economically viable at scale**. At Coforge, I re-architected inference systems from blocking to async, scaling from ~20 to 140+ calls per VM, supporting 1,600+ concurrent sessions, and driving ~$1.3M in annualized compute savings. I’ve reduced MTTR from hours to minutes by building observability into the system from day one—not as an afterthought. Across roles, I’ve built: * Real-time conversational AI systems (Voice → STT → LLM) under strict latency constraints * Retrieval-augmented systems over large enterprise corpora with grounded responses * Low-latency recommenders and multilingual AI systems operating at scale (163+ languages) * End-to-end AI platforms—from model orchestration to production deployment I’m particularly interested in: * Inference systems and low-latency architectures * Retrieval systems (RAG) and failure modes in LLM pipelines * Observability, reliability, and cost optimization for AI systems I don’t focus on building demos. I focus on building systems that survive production.

Location

Bengaluru

AI stack

claude code

codex

ollama

Work experience

SWE 1 at Skan AI, Jun 2026 - Present · 4 mo, Bengaluru, Owning the end-to-end lifecycle of process engineering and synthetic data generation across Skan’s Agentic Products stack. Translating client clickstreams into process definitions, core SOPs, and BPMN workflows, while independently designing novel business processes and human-like execution patterns. Building supporting web applications and simulation platforms, engineering Windows-level automation using Playwright, Win32 APIs, and ctypes, and auditing agentic VA event capture and AOP (Agentic Operating Procedure) pipeline outputs to ensure behavioral fidelity, process correctness, and end-to-end system reliability.

AI Engineer at Coforge, Jun 2024 - Jun 2026 · 2 yrs 1 mo, Bengaluru, > Built and scaled production-grade conversational AI systems (ASR → LLM → TTS) for enterprise and banking use cases, handling real-time, high-concurrency workloads > Re-architected inference stack to overcome python GIL and migrated threaded workflows to async (`asyncio + uvloop`), improving throughput from ~20 → 140+ calls per VM (~7x gain) > Scaled infrastructure to 1,600+ concurrent sessions with <2s performance under production load > Reduced MTTR from ~1–2 hours to ~5 minutes by implementing structured logging and observability pipelines > Delivered ~$1.3M annualized compute savings through infra optimization (VM right-sizing, concurrency tuning, 32-core → 8-core strategy) > Built LLM-powered systems across domains: banking (real-time voice automation), airlines (contract intelligence), and Azure (infra documentation + knowledge retrieval) > Led LLM adoption internally by training 130+ engineers on Java Spring AI > Recognized with “Best Team” award and "Pat on the back - Think Customer" award

Data Scientist at Gida Technologies, Jan 2023 - May 2024 · 1 yr 5 mos, Bengaluru, > Built and productionized three interlinked AI platforms spanning multilingual content generation, chatbot builders, and API toolkits. > Designed a graph-based skill recommender system achieving ~90% accuracy with sub-second inference on GPU-backed infrastructure > Developed a RAG-based multilingual chatbot pipeline (~97% factual accuracy), enabling grounded, scalable interactions across diverse corpora, supporting 163+ languages at scale > Optimized systems for low-latency inference and high-throughput workloads, including sub-50ms response times for recommendation pipelines > Contributed to end-to-end AI product delivery, from model design and retrieval pipelines to deployment-ready APIs and system integration

Head of Machine Learning at Indian Institute of Science (IISc), Jan 2022 - Sep 2022 · 9 mos, Bangalore Urban, Karnataka, India, Led and established the Machine Learning Team at NMCAD Lab, Aerospace Engg. Dept., IISc which is currently working on the design optimisation task using ML and DL for their eVTOL design under Dr Dineshkumar Harursampath, guided by Dr Ramesh Gupta & Harsh Tiwari. Being the founding leader of the machine learning team at IISc, I was posed with the task of birthing an entire new vertical in the Phase - II of the project. From forming the team to leading it to being the top one in a span of 8 months with 5 vital projects, I was tested for a wide array of technical, managerial and leadership proficiencies.

AI Product Developer at CellStrat, Feb 2021 - Dec 2021 · 11 mos, Bangalore Urban, Karnataka, India, Built readily usable and functional AI products based on OpenAI's research publications for the consumer base of their in house AI-as-a-service platform, cellstrathub.com which hosted over 11k global AI Developers at the time of projects' deployment alongside a few other projects handed over to test the possibility of launch of new verticals at CellStrat.

Education

Executive Diploma, Artificial Intelligence and Machine Learning, International Institute of Information Technology Bangalore

Bachelor of Engineering - BE, Mechanical Engineering, B. M. S. College of Engineering

Public links

github: https://github.com/spice14

ashwingupta.dev: https://ashwingupta.dev/