Headline

Agentic AI Engineer @ Reliance Intelligence | LLM Agent Orchestration · Multilingual Voice AI (Indic) · RAG · LLMOps | Ex-Amazon · Walmart | Open to relocation

About

I build production agentic AI systems: the orchestration harness, tool calling, failure recovery, memory and evals that make LLM agents fast and reliable for real users. Most recently I am on the agentic orchestration team behind BharatIQ, Jio’s multimodal, multilingual AI assistant (friends-and-family beta, tens of thousands of users). There I built a deterministic fast path that cut tool-triggered responses from 20–90s to ~1s. I also built fault-tolerant turn recovery so text and voice answers are never silent, dual-channel voice streaming for code-mixed Hindi–English (Hinglish), agent evals using golden sets and LLM-as-judge, and per-query LLM cost attribution and quotas. This sits on 7 years of software engineering at Amazon, Dremio (upstream Apache Arrow / Gandiva commits), Walmart Global Tech and a founding-team startup, covering distributed systems, workflow orchestration (Temporal) and query engines. At Walmart I shipped my first production LLM system, a RAG on-call assistant that auto-diagnosed ~85% of frequent alerts. What I work on: • Agent orchestration & custom harness design · tool/function calling · MCP · semantic skill routing • Multimodal & multilingual voice agents · TTS/SSML · ASR · Indic languages & code-mixing • RAG · embeddings · Qdrant · Pinecone · re-ranking · guardrails • Agent evals · golden datasets · LLM-as-a-judge · model benchmarking (GPT-5, Gemini, Claude) • LLMOps · LLM gateways · cost attribution · latency (P50–P99, TTFT) · OpenTelemetry · Langfuse • Python · FastAPI · LangGraph · LangChain · Google ADK · GCP · AWS · Kubernetes Open to Agentic AI Engineer, AI Engineer and LLM Engineer roles (full-time, contract or part-time) in India, Europe and beyond. Open to relocation. Contact: sagnik31@gmail.com · github.com/sagnik31

Location

Bengaluru

AI stack

claude code

cursor

hugging face

mcp

Google ADK

Qdrant

Work experience

AI/ML Engineer at Reliance Intelligence, Mar 2026 - Present · 9 mos, Bengaluru, Karnataka, India, JioBharatIQ - Jio’s multimodal, multilingual AI assistant (pre-launch; friends-and-family beta with tens of thousands of users). Consultant on the agentic orchestration team, building core components of the custom agent harness. • Deterministic fast path: app-action taps and predefined triggers call the bound tool directly with zero LLM inference, cutting those responses from 20–90s to ~1s; circuit-breaker fallback to the full agent loop; feature-flagged, cross-team rollout. • Fault-tolerant turn recovery: forced fallbacks on failed tool calls, degraded-mode answers on timeouts, a never-silent voice safety net, and async background completion with an “answer ready” push when users leave mid-response. • Multilingual voice: one streamed LLM response feeds both spoken TTS and the on-screen answer, with provider-specific SSML (ElevenLabs, Google Chirp) for Hindi, English and Hinglish. • Agent evals & benchmarking: golden datasets and LLM-as-judge scoring response quality plus skill-routing and tool-calling accuracy over a Qdrant skill registry; GPT-5 vs Gemini (Vertex AI) latency and quality benchmarks that informed the production model migration. • LLMOps: per-query cost attribution through the LLM gateway (fixed ~2.2x token under-counting on tool turns), per-user quota design, ~100x smaller maps-tool payload (271K → 3K tokens), provisioned-throughput capacity audit. • Observability & releases: OpenTelemetry, New Relic distributed tracing and AI monitoring, Langfuse tracing fixes; owned release trains v1.13–v1.18. • Prototypes: Mem0 long-term memory, MCP memory tools, Google ADK computer-use agent, AG-UI/A2UI generative UI streaming.

Freelance AI & Backend Engineer at Stealth Startup, Jul 2025 - Mar 2026 · 9 mos, Remote, • Jatayu Healthcare Technology (VoiceDocAI): architected B2B API-key licensing, session-token rotation, and real-time concurrency and daily quota enforcement for an LLM-powered doctor–patient transcription product, enabling third-party HMIS integrations designed for ~1,000 concurrent users. • Geospatial analytics: spatial population-density pipelines (Python, GeoPandas, Uber H3) for radiology catchment analysis. • Applied AI R&D: local VLM/SLM RAG prototype with retrieval-precision and hallucination evals; agentic workflow design and self-hosted LLM infrastructure.

Software Engineer III at Walmart Global Tech, Dec 2023 - Jun 2025 · 1 yr 7 mos, Bengaluru, Karnataka, India, Seller Risk platform, Walmart Global Marketplace (fraud detection for high-risk sellers at scale). • Built an LLM-powered on-call assistant end to end: a RAG pipeline (Pinecone, retrieval evaluation, layered guardrails) in Slack that auto-diagnosed ~85% of frequent alerts and cut manual triage ~40%. • Engineered high-throughput risk-detection workflows over Marketplace seller data, publishing real-time fraud signals to Salesforce. • Built a BFF aggregation layer consolidating downstream risk microservices into a single API for real-time internal dashboards.

Senior Software Engineer at BeepKart, Feb 2022 - Sep 2023 · 1 yr 8 mos, Bengaluru, Karnataka, India, Part of the founding team, digitising the secondhand two-wheeler market in India. • Shipped 3 major Java microservices from scratch, replacing legacy Magento/PHP code; improved lead generation volume by 50%. • Implemented an end-to-end vehicle refurbishment workflow using Temporal (workflow orchestration), increasing throughput by 200%.

Software Engineer at Dremio, 2019 - 2021 · 2 yrs, Hyderabad, Telangana, India, Worked on the query execution team at Dremio, a data lake engine built on Apache Arrow. • Committed upstream optimisations to open-source Apache Arrow (Gandiva), improving performance of core SQL functions. • Implemented Parquet record readers and various SQL execution operators for cloud data lake sources.

Education

Bachelor of Engineering (B.E.), Computer Engineering, Jadavpur University

Higher Secondary, Science, South Point High School, Kolkata