deepaksharma

Lead AI / Staff ML Engineer · Bengaluru

Deepak Sharma.

I build LLM systems that ship: multi-agent workflows, human-in-the-loop review, long-running agents, and the retrieval and infrastructure under them.

8+ years taking ML from notebooks to production. Founding ML engineer at a sales intelligence platform that grew from 0 to $10M ARR, where I owned the AI architecture end to end.

Today I build agentic document extraction with human-in-the-loop review at Zenskar.

Get in touch

Work.

2017 — now

  1. Senior AI/ML Engineer, Zenskar

    • Document-extraction agent that pulls structured billing data from PDF invoices and contracts. Every field carries a citation to its source, and low-confidence fields are routed to a human-in-the-loop review step before they reach billing.
    • Entity resolution agent step that reconciles extracted entities against existing billing records, escalating ambiguous matches to a reviewer.
    • Internal knowledge assistant built as a tool-using agent: it navigates the knowledge base with bash, grep and file search (agentic retrieval) instead of an embedding pipeline.

  2. Founding ML Engineer, OrbitShift

    • 0 → $10M ARR. Owned AI architecture end to end: data pipelines, retrieval, serving and monitoring.
    • Hybrid RAG on Elasticsearch (keyword + vector); embedding work lifted retrieval accuracy 6% in production.
    • Agents with memory for RFP automation and sales outreach, holding context across multi-turn, multi-step sessions.
    • LLM platform with agent orchestration, RBAC, cost tracking and token monitoring; eval harnesses and drift monitoring before and after deploy.
    • Hired and led a team of 3 ML engineers.

  3. Data Scientist, ConcertAI

    • Predictive models on multi-source clinical data, with MLflow tracking and SHAP explainability.

  4. Senior Research Analyst, Innoplexus

    • NLP pipeline for biomarker discovery from literature; gene–gene network database for pathway research.

  5. Bioinformatics Engineer, Elucidata

    • Python tooling for biomedical data analysis with research labs and pharma teams.

What I work on.

01 · Multi-agent systems

Orchestrator and worker architectures, tool calling, handoffs and shared memory, built to be debuggable.

02 · Human-in-the-loop

Confidence-based routing, review queues, citation grounding and feedback loops that turn corrections into evals.

03 · Long-running agents

Durable execution, checkpointing, retries and resumable state for tasks that run for minutes or hours.

04 · Fine-tuning

LoRA / QLoRA and PEFT: dataset curation, training, and evaluation against a prompted baseline.

05 · Retrieval

Hybrid search, embedding tuning, reranking and agentic retrieval over tools and files.

06 · Evaluation

Eval harnesses, LLM-as-judge, tracing and drift monitoring before and after deploy.

07 · LLM platforms

Orchestration, RBAC, guardrails, cost and token governance across teams.

08 · Agent tooling

MCP servers and tool design so agents can act safely on real systems.

Stack.

PythonPyTorchLoRA / PEFTMulti-agent orchestrationMCPDurable workflowsLLM evalsFastAPIElasticsearchVector search AWSDockerMLflowPySparkscikit-learnSHAP