I build production AI systems where the model is the product.

AI/ML engineer with 4+ years shipping LLM systems to real users: a speech-AI service for a US telephony provider, the multi-agent engine that lets a 12B model do 32B-class work, and RAG over 10M+ embeddings, often on-prem and air-gapped.

Download CV mohdsalimkt@gmail.com LinkedIn GitHub

12B → 32Bclass work from a smaller model

10M+embeddings in production RAG

50k+concurrent users on live captions

8Kubernetes environments incl. GPU/DGX

Flagship: Echo

Production speech AI for a US hosted-telephony provider's voicemail and call-recording features.

Speech AI · in production

Two flows, two engines, one reliable service

Voicemail, accuracy-first: NVIDIA Canary, with automatic fallback to Whisper when output is degenerate, and a final failover to the shared vLLM tier.
Call recordings, throughput-first: Whisper served on vLLM behind a load balancer, Redis priority queues (0–99) with live re-prioritisation, and Celery workers on stateless HA nodes.
Long audio: 45-minute chunks up to 10 hours, with timestamps kept absolute across chunks; automatic language ID.
Speaker diarization: Silero VAD per channel for stereo; ECAPA-TDNN embeddings with agglomerative clustering for mono.
LLM analytics that never block transcription: Gemma summaries and keywords on a separate queue and load balancer, delivered by webhook or polling.
Operations: hash-based duplicate suppression, per-tenant correction rules, CUDA-error watchdog, fork-safe PyTorch in Celery.

Lesson I carry everywhere: a 200 OK isn't success. Validate the model's output, not the status code, and fail over automatically.

NVIDIA CanaryWhispervLLMFastAPICeleryRedisCassandraSpeechBrainSilero VADGemma

Echo architecture: voicemail flow with Canary to Whisper to vLLM failover, call-recording flow with API, Redis priority queue, Celery workers and Whisper on vLLM, and a separate analytics queue with Gemma — Echo architecture (host and customer names omitted)

Selected work

Everything here runs in production or is public on GitHub.

Agents

Multi-agent tool-calling engine

Replaced CrewAI · 12B model doing 32B-class work · GPU cluster → single node

I wrote the engine behind an enterprise on-prem LLM platform (~49k lines, largest contributor): a 73-toolset registry with schema-validated tool calls, ReAct with tiered escalation, structured delegate / ask / finish steps, a FAST/SMALL/LARGE/THINKING model router over vLLM and Ollama, SSE event streaming and Redis-backed admission control.

Agent engine layers: clients, admission and routing, orchestrator, tools, model router, quality

PythonFastAPIvLLMOllamaRedisKubernetes

Retrieval

Enterprise RAG platform

10M+ embeddings · 10,000+ queries/day · on-prem & air-gapped

Ingestion with Docling for layout-aware parsing (PDFs, DOCX, tables, figures) and RapidOCR / PaddleOCR-VL / Qwen-VL for scanned documents; structure-aware chunking; hybrid dense + keyword retrieval with re-ranking. Runs on Cassandra Vector, pgvector, Qdrant and FAISS across 8 Kubernetes environments, built to HIPAA-style controls.

DoclingRapidOCRpgvectorQdrantFAISSSentence-Transformers

Speech & Vision

Live captions and document AI

92% caption accuracy · 50,000+ concurrent WebSocket users

Whisper speech-to-text and Tacotron2 text-to-speech shipped into a video-calling product. Document AI with PaddleOCR-VL and Qwen-VL served via Transformers, LayoutLMv3 fine-tuning and multiple OCR engines feeding the RAG pipeline.

WhisperTacotron2WebSocketsPaddleOCR-VLQwen-VLLayoutLMv3

Arabic AI

Arabic LLM fine-tuning

Arabic-language models · LoRA / QLoRA

LoRA / QLoRA fine-tuning of open models for Arabic with Unsloth and PEFT, plus Arabic-aware retrieval (normalisation and hybrid search). I am proficient in reading and understanding Arabic, which helps when checking model output for Gulf products.

UnslothLoRAPEFTHugging Face

Automation

AI ticket triage with human approval (n8n)

LLM drafts · coded guardrails decide · a human approves risky cases

A webhook takes a customer ticket; an LLM classifies it and drafts a reply as structured JSON; deterministic rules route refunds, warranty, low-confidence or high-value cases to an Approve / Edit / Reject form. Tested on 8 scenarios including Arabic and a prompt-injection attempt.

n8n workflow from webhook through LLM classification and guardrails to auto-reply or human approval

n8nOpenAIStructured outputsHuman-in-the-loop

Open source

small-model-agents

Agent evaluation harness · CI regression gate · MCP server

Public work on making small models reliable agents: an evaluation harness with a CI gate that blocks regressions, plus a published Model Context Protocol server (stdio JSON-RPC).

github.com/salimkt/small-model-agents →

EvalsMCPGitHub Actions

How I build AI products

Three principles from shipping systems that real people depend on.

Treat model output as untrusted

Structured outputs, validation and guardrails in code, and a human in the loop for high-risk actions. Rules decide; the model reads and writes.

Design for cost and failure

Return fast, queue behind, put model tiers behind load balancers with failover, and route to small models first, escalating only when needed.

Measure everything

Every change ships with an eval set and a CI regression gate. I track recall, latency and cost per request, so quality is provable, not anecdotal.

Skills

Hands-on, in production unless noted.

LLMs, agents & RAG

Tool callingMulti-agentCrewAILangChainLlamaIndexMCPHybrid searchRe-rankingEvalsPrompt engineering

Models & serving

vLLMOllamaTransformersPyTorchUnslothLoRA / QLoRAQuantizationWhisperNVIDIA Canary

Data & vectors

pgvectorQdrantFAISSCassandraPostgreSQLMongoDBRedisDoclingOCR

Backend

PythonFastAPIDjangoCeleryWebSocketsSSENode.jsREST

Platform

KubernetesDockerGPU / DGXAir-gapped deploysGitLab CIPrometheusGrafana

Apps & integrations

ReactReact Nativen8nWhatsApp Business APISIP / VoIP

Experience & education

Jun 2022 – present

AI/ML Engineer · Netstratum Technologies, Kochi

Production LLM, speech and document-AI systems for an enterprise on-prem AI platform and its customers: Echo, the agent engine, RAG, vision, and Kubernetes delivery across 8 environments.

2018 – 2022

B.Tech, Computer Science & Engineering

KMEA Engineering College, APJ Abdul Kalam Technological University.

Languages

English (fluent) · Malayalam (native)

Also Hindi and Urdu; proficient in Arabic (reading, comprehension and writing).

Let's build your AI core.

Happy to talk about AI engineering roles, consulting and collaborations.

Email me LinkedIn CV (PDF)