Software Engineer (AI/ML): LLM systems, RAG, agent orchestration, speech & vision
4+ years shipping production AI · Kochi, India
Portfolio · LinkedIn · mohdsalimkt@gmail.com
I build production LLM systems end to end: agent orchestration, retrieval, speech pipelines, model serving, and the evaluation that keeps them honest. Most of it runs on-premise and air-gapped inside enterprise customers' own infrastructure.
- 12B → 32B-class: my multi-agent engine lets a 12B model do 32B-class work, moving each deployment from a GPU cluster to a single node
- 10M+ embeddings in production RAG, serving 10k+ queries/day
- +45% answer accuracy over the bare-LLM baseline, 1.8 s average inference latency, 99.7% uptime
- 94% speech-to-text accuracy and 92% live-caption accuracy in production
- 50k+ concurrent users on real-time WebSocket / REST services
- 8 Kubernetes environments, including two GPU/DGX clusters
Multi-agent tool-calling engine. bm-agent-engine replaced the CrewAI-based orchestration of an on-premise enterprise LLM platform and now serves all its agent traffic; I was the largest contributor to its ~49k lines. A 73-toolset registry, ReAct agents with tiered escalation, and a FAST / SMALL / LARGE / THINKING router over vLLM and Ollama. Because the SSE event contract was kept exactly, the UI cut over with a single config change.
Enterprise RAG. Docling + RapidOCR ingestion, Sentence-Transformers embeddings, a Cassandra vector schema, and hybrid keyword + semantic retrieval with re-ranking over 10M+ embeddings. Quantization, batching and caching brought average inference to 1.8 s; model failover and warm caching hold 99.7% uptime.
Speech AI in production. A transcription service for a US hosted-telephony provider's voicemail and call recordings: NVIDIA Canary and Whisper on vLLM with automatic failover, Redis priority queues and Celery workers, recordings up to 10 hours, and speaker diarization. Also: Whisper STT (94%), Tacotron2 TTS and live captions (92%) in a video-calling product.
Vision and documents. PaddleOCR-VL and Qwen-VL served through Transformers, a pluggable OCR layer feeding RAG, and LayoutLMv3 fine-tuned to 91% F1 on 100K+ documents a month.
| Project | What it is |
|---|---|
| DataRat · live demo | AI data analyst. Connect 11 kinds of data source and get KPIs, anomalies and Holt-Winters forecasts; a Planner → SQL → Analyst → Narrator agent pipeline answers plain-English questions, with a rules engine as fallback and a read-only SQL guard. FastAPI, DuckDB, React. |
| 3xDezine · live demo | Home design and costing. 2D floor plan to WebGPU 3D walkthrough with a live material estimate and GST-ready quotation; 300+ unit tests and 136 pgTAP database assertions. TypeScript, Fastify, Supabase, Kotlin. |
| small-model-agents | Reference implementation of reliable agents on small models: schema-validated tool calling, tiered routing, evidence-based escalation, an MCP server, and an eval harness whose CI gate blocks regressions. Python stdlib. |
| web-share · live demo | Peer-to-peer WebRTC file sharing as a PWA, with end-to-end-encrypted MQTT signalling or offline QR pairing. No server, no accounts. |
| Project | Contribution | Status |
|---|---|---|
| Celery | Log at INFO when a prefork child is recycled | In review |
| PaddleOCR | MCP server and doc2md fixes: #18416, #18417, #18418 | In review |
| vLLM | Design for multi-channel audio reduction | Proposed |
| silero-vad | Asymmetric speech padding | Proposed |
Python FastAPI Django Celery vLLM Ollama Hugging Face PyTorch LangChain CrewAI MCP
Cassandra pgvector Qdrant FAISS PostgreSQL DuckDB Redis Kafka RabbitMQ
Whisper NVIDIA Canary Tacotron2 LayoutLMv3 PaddleOCR-VL Qwen-VL
Kubernetes Docker GCP GitLab CI GitHub Actions
TypeScript React React Native three.js Kotlin
Software engineer (AI/ML) at Netstratum Technologies since June 2022. Contact: mohdsalimkt@gmail.com




