I'm an AI Engineer specializing in Agentic Workflows architecting autonomous systems that plan and self-correct. Currently, I am engineering a SaaS-grade geospatial lead generation engine alongside a GenAI marketing workflow using Gemini and Nano Banana Pro to automate high-fidelity Meta ad creation.

My focus is on Human-in-the-Loop pipelines that turn probabilistic models into deterministic business value. Leveraging Python, LangGraph, FastAPI, and Docker, I build robust architectures designed to solve production challenges like inference latency and RAG hallucinations.

From February to September 2026 I was the only backend engineer on Saroya AI, a mobile app from TinyCheque (an AI-first venture studio) where creators build an AI avatar that their audience chats with. I wrote the API from scratch, built the ingestion pipeline, and ran the service from May 2026 until it reached testers on the Play Store and TestFlight. I write the eval set before I trust a change, so the numbers below come from a 71-question golden set that runs in CI.

Graph-database queries were slow enough to time out streaming responses, so I moved retrieval to Postgres with pgvector and full-text search, fused with reciprocal rank fusion (RRF), and added mem0 for long-term memory.

flowchart LR
    U[User message] --> V[pgvector search]
    U --> F[Full-text search]
    V --> R[RRF fusion]
    F --> R
    M[mem0 memory] --> P[Prompt and cache]
    R --> P
    P --> L[OpenRouter<br/>5 providers, fallback]
    L --> S[SSE stream to mobile app]

_{Simplified view of the chat path.}

• 🛠️ AI Engineer with production experience in building Agentic RAG systems, and GenAI Marketing Pipelines.
• 🧠 Hands-on experience training and refining LLMs, developing CNN architectures, and model deployment.
• 🔧 Proven track record of delivering business impact, including reducing content production costs by 100% via automated GenAI workflows (Gemini/Nano Banana)
• 🚀 Sole backend engineer on a production GenAI mobile app: retrieval recall up from 0.53 to 0.98 and latency down from 5 to 8 s to 0.6 s, verified by an eval harness and gated on SLOs in CI.
• 💸 Built the LLM cost controls for that app, including prompt and reply caching and provider fallback, which reached a 78% cached-input ratio.

Feb 2026 to Sep 2026

• Retrieval: Cut retrieval latency from 5 to 8 s to 0.6 s and lifted recall from 0.53 to 0.98 (MRR 0.89). I moved retrieval to Postgres pgvector with full-text search, RRF fusion and mem0 memory, then wrote a 71-question golden set and eval harness to check that quality held.
• Backend: Built and ran the whole backend alone. I wrote the TypeScript/Node/Express API from scratch (auth, Razorpay payments, creator onboarding, chat) and operated it from May 2026 through the release to testers on Play Store and TestFlight.
• LLM cost: Real routing costs overran my cost model, so I instrumented every call, added prompt and reply caching, and pinned OpenRouter to 5 providers with fallback. The cached-input ratio reached 78%, and a weekly CI check alerts below 70%. Tokens stream to mobile over SSE with pacing and keepalive.
• Ingestion: Built the pipeline behind every avatar. It pulls creator content from X, Reddit, YouTube and LinkedIn, plus PDF, DOCX, XLSX and PPTX uploads with OCR. I added KYC identity checks and hard-delete workflows across every data store.
• Pricing: Originated the creator pricing and credit model the CPO adopted, plus the PRDs, ADRs and cost analyses the team planned against.
• Tooling: Set up a Claude Code harness with 14 subagents, 22 slash commands, and MCP tools and hooks that enforce spec-first and test-first work, backed by a 169-test suite.

Remote | Jan 2025 – Feb 2026

• GenAI Pipeline: Architected a marketing workflow using Nano Banana Pro (Gemini) and LangChain to generate photorealistic product assets, cutting photoshoot costs by 100%.
• SaaS-Grade Tooling: Engineered a Geospatial Lead Finder using Google Maps API that extracts global B2B leads with sub-second latency.
• Deployment: Built a Streamlit dashboard allowing non-technical staff to trigger complex scraping and generation tasks.

Remote | Mar 2025 – Aug 2025

• RLHF & Data Quality: Optimized AI-generated code by correcting logic errors to create high-quality ground truth data for training foundation models.
• Safety & Compliance: Enforced strict safety guidelines on code outputs to ensure compliance for production deployment.

Remote | Nov 2024 – Dec 2024

• Custom CNNs: Architected a multi-class object recognition model (vehicles, biological entities), achieving 90% accuracy via ensemble techniques.
• Real-Time Inference: Deployed a live inference pipeline via Streamlit for instant visual validation.

I don't just train models, I break them to see how they fail.

• Agent First: I believe the future isn't bigger models, but smarter agentic loops. I code for the future where AI plans, not just chats.
• Rapid Learner: Learned Swarm Robotics from scratch in 3 weeks for my dissertation because standard robotics felt too "safe."

• Current Focus: Architecting "Human-in-the-Loop" systems for B2B automation.
• My Philosophy: An AI model is only as good as the pipeline around it. I spend 20% of my time on the model and 80% on the orchestration (LangGraph), data engineering (RLHF), and deployment (FastAPI/Docker).
• What I'm Building: A geospatial lead-generation engine that combines Google Maps API with autonomous scraping agents to find B2B leads globally.
• Language Mix: Python first, plus TypeScript for the production backend I shipped this year.

My recent production work is in private repos, so these are the public ones.

BSc (Hons) Artificial Intelligence, 2:1, University of Manchester (2021 to 2024). Dissertation: simulation and optimization of the beta algorithm for autonomous swarm robotics.

• Agent Harness:

• Languages:

• AI/ML:

• Tools:

• APIs:

• Backend & DevOps

• Also in use: Express, pgvector, FalkorDB, ChromaDB, OpenRouter, mem0, Cloudflare R2, Apify, Railway, OpenTelemetry, SigNoz, Claude Code (custom subagents, skills, hooks), MCP

I am actively looking for AI Engineering roles where I can deploy agentic systems to production. I'm based in Kolkata and available immediately, and I'm open to applied AI, forward-deployed and AI solutions engineering roles in India, Germany and the UK.