Post Job Free
Sign in

Deployed AI Solutions & Agentic Engineer

Location:
New Delhi, Delhi, India
Salary:
20 LPA
Posted:
October 25, 2026

Contact this candidate

Resume:

Forward Mann Deployed Engineer Vishnoi · Solutions Engineer

+91-837******* **************@*****.*** in/mann-vishnoi-814438181 github.com/mannanoyed Delhi NCR · relocation & remote Summary AI engineer who deploys into customer environments. Production LLM infrastructure since 2024: multi-agent orchestration, MCP, RAG, and the gateway metering every model call for a 53-person organisation. Then out to enterprise clients in the UAE across steel fabrication, ceramics and automotive retail: runs the discovery calls, authors the solution architecture, builds the POC on their own production data, and presents it to their engineering reviewers. No handoffs. Work Experience

ATGWorld Networks Pvt Ltd (Banao Technologies) Apr 2024 – Present Technical Lead, Applied & Agentic AI Nov 2024 – Present AI Engineering Intern, Product Systems Apr – Oct 2024 · Hybrid

• Lead the applied-AI function across all three products: architecture, release planning, and a 30-person intern cohort with two full-time engineers.

Client Engagements — UAE · Banao Technologies, ATG services arm

• Assent Steel (Dubai; structural steel): Rs.1–1.5M/yr recoverable, scaling to Rs.5–7.5M at full platform, computed from the client’s own revision data; one revision alone surfaced Rs.49,020 of reusable stock. Passed engineering review; scope expanded to a cutting-stock optimiser over ~2,300 tons. Designed the Material Optimization Platform behind it: GUID-keyed line items diff revisions deterministically into a cross-project excess pool, so one project’s redundant steel becomes another’s supply.

• RAK Ceramics (ADX-listed, ~$1B turnover): POC on a bounded inverse of the manufacturing response, specified to remove 5 of every 8 print-fire-measure cycles: shade approval from 7–8 iterations to 2–3, each avoided cycle a full kiln firing. Maps measured CIELAB and gloss deviation into CMYK corrections calibrated per SKU and kiln, with an operator gate because a wrong correction fires a whole batch.

• Genesis Motors (Dubai): Every showroom walk-in scored on an intent ladder and written into CRM as structured lead intelligence. Built a bilingual English/Arabic voice concierge: Whisper to GPT-4o over ChromaDB RAG to streaming TTS. AXON — company-wide work intelligence platform

• Architected AXON, a five-service platform that turns scattered work signal (Google Meet calls, Slack threads, the ERP, the intranet) into one live graph of who owns what and which commitments are slipping, across a 53-person organisation. Every exchange becomes an event: IRIS perceives, NERVE routes, CELL executes, CORTEX reasons, STROMA sustains.

• No duplicate ERP writes across restarts or provider outages: the event bus is a transactional outbox on PostgreSQL with SKIP LOCKED delivery, backoff to 1h, dead-lettering and event-ID idempotency. Root-caused an org-wide outage to one jsonb encoding bug.

• 99.1% of a repeated prefix served from cache, org-wide, because prompt caching belongs in GATE, the gateway metering every model call. Replaced a racy budget precheck with a two-phase reserve/settle ledger. axon-code, the coding agent above it, runs on a four-invariant permission ladder: un-opted repos read-only, autonomy policy from the protected branch, and a Rule of Two denying untrusted input combined with external reach and mutation. InterviewGod — AI hiring platform: interview, assessment, screening, proctoring, evaluation

• Founding AI engineer. Authored the evaluation engine behind every candidate score, in production ten months on. Each answer graded twice: against a generated ideal reference answer and a holistic transcript pass. Its retry layer refuses parse errors, so a malformed-prompt bug cannot hide as flakiness. Also added a filterable signal profile, weekly rounds measuring improvement drift, a vector-retrieval question bank with ranking, and voice-comparison cheat detection. Vidhyaa — multi-agent learning platform

• Every learner gets a walkthrough generated for them rather than a shared recording. A multi-agent tutor where agents plan the lesson, generate slides, and drive a code-rendering node that types and records the code live. Skills

LLM & Agents: Anthropic Claude API, OpenAI API, Gemini, LangChain, LangGraph, MCP (Model Context Protocol), RAG, vector databases

(pgvector, Pinecone, ChromaDB), embeddings, tool calling, multi-agent orchestration, coding agents (Claude Code, Cursor), prompt engineering, prompt caching, guardrails, LLM evaluation Backend & Data: FastAPI, async Python, PostgreSQL, MongoDB, SQLAlchemy, Alembic, REST, SSE, webhooks, event-driven architecture, transactional outbox, idempotency, circuit breakers, distributed systems, pytest Cloud, DevOps & MLOps: AWS (Lambda, EC2, S3, SQS, SageMaker), Docker, GitHub Actions CI/CD, Linux, nginx, Cloudflare R2, RunPod GPU, Git, observability, cost governance, incident response, MLOps, LLMOps Languages: Python, TypeScript, SQL, C++, R, Bash

Machine Learning & CV: PyTorch, TensorFlow, Hugging Face, OpenCV, Whisper, speaker diarization, reinforcement learning (DQN) Education

Vellore Institute of Technology, Master of Computer Applications (MCA) Oct 2022 – Jul 2024 CGPA 8.52 / 10 · Best Research Paper award for expressive speech synthesis (Tacotron with emotion embeddings)



Contact this candidate