Sathish Kumar V
Open to UK/US remoteRésumé

AI Solutions Architect · Team Lead · 7+ Years · Chennai → UK/US Remote

SathishKumar.

I architect agentic AI for the enterprise — autonomous systems with memory, reasoning loops, tool orchestration, and the guardrails that make them safe to ship. From MCP servers to RAG pipelines, I build AI that acts, not just responds.

See selected workView résumé

99%

Time reduction

Contract extraction: 260h → 26min

15

AI products owned

Enterprise platform, agents, MCP tools

7

Engineers led

Hiring, mentoring, delivery

5

Deals won

Solutions I architected and pitched

01Expertise

What I architect.

Six domains where I design, build, and ship production systems.

Agentic systems

01

Autonomous agents with reasoning loops, tool orchestration, and human checkpoints — architected to fail gracefully.

  • LangGraph orchestration
  • MCP tool harness
  • Memory & context management

RAG & knowledge

02

Retrieval systems that cite sources, surface uncertainty, and know when to escalate to humans.

  • Hybrid vector + keyword search
  • Re-ranking pipelines
  • Citation & provenance tracking

MCP & tool integration

03

Secure, scoped MCP servers that let agents act on enterprise systems with full audit trails.

  • Custom MCP protocol servers
  • Microsoft Fabric integration
  • Governed API access

Document intelligence

04

OCR pipelines handling complex layouts — nested tables, merged cells, multi-page contracts — with confidence-based routing.

  • PaddleOCR + LLM extraction
  • Table structure detection
  • Human-in-the-loop routing

Data foundations

05

Semantic layers and governed access so AI answers match the numbers business users trust.

  • dbt transformation models
  • Power BI semantic layer
  • Data contracts & lineage

Governance & observability

06

Guardrails, audit trails, and evaluation frameworks that risk teams can sign off on.

  • RBAC & access control
  • LLMOps & Langfuse
  • Compliance documentation

02Selected work

Systems that shipped.

Client delivery and open source — each with the architecture decisions and production lessons learned.

/01Document Intelligence

Contract Extraction at Scale

Enterprise clients drowning in 200-page contracts with pricing buried in nested tables. Manual extraction took 260 hours per batch; commercial OCR failed on complex layouts. I architected a hybrid pipeline combining PaddleOCR for structure detection with Qwen 7B for field extraction.

Outcome85% straight-through automation. Cost reduced from $0.03/page (GPT-4V) to $0.002/page. Four-week PoC became six-month production engagement.
PaddleOCRQwen 7BHuman-in-loopSalesforce
/02MCP · Open Source

Custom Fabric MCP Server

MCP server enabling AI agents to query Microsoft Fabric with validation, Excel export, and dashboard generation. Extends native capabilities with governed access patterns suitable for enterprise deployment.

MCP ProtocolPythonMicrosoft FabricAzure
/03Enterprise · Platform

Private AI for 200 Consultants

Consultants using public ChatGPT with client data created compliance risk. Built a governed internal platform connecting SharePoint, Power BI, and CRM with RBAC — same AI capabilities, zero data leaving the perimeter.

Azure OpenAIRBACKnowledge graph
/04RAG · Knowledge

AI Context Layer for PE

Investment teams spent days preparing for committee meetings. Built an AI layer over deal rooms, financial models, and reports with citation tracking, assumption challenging, and full audit trail.

Claude EnterpriseMCP ProtocolCitation tracking
/05AI · Open Source

RMG: Resource Management System

Replaced manual email-based resource allocation with AI-powered candidate ranking. Six-step pipeline: semantic matching, weighted scoring (skills 40%, competency 25%, availability 25%), GPT-4o enrichment.

PythonGPT-4opgvectorFastAPI
/06Data · Platform

Data Engine for PE

High-performance data engineering platform for Private Equity portfolio companies. Centralized command center managing full data lifecycle with medallion architecture (Bronze → Silver → Gold).

TypeScriptdbtRedshiftPower BI
/07Agents · Infrastructure

SK-Hermes: Always-On Agent Runtime

Hermes Agent fork with documented AWS EC2 Docker Compose deployment. Personal always-on runtime for autonomous agent workflows and experimentation.

PythonDockerAWS EC2Agentic AI

03How it runs

Watch it think.

The agentic architecture pattern I deploy, with a real decision trace from production.

Reference architecture · agentic enterprise platformproduction · v3
Experience
Web chatTeams botAPI endpoints
Agent harness
Planning loopAgent routerMemory storeHuman checkpoints
Tool layer · MCP
CRM serverFabric serverDocs serverSharePoint connector
Knowledge & retrieval
Vector indexSemantic modelsDocument storeGraph DB
Model layer
Azure OpenAIClaudePaddleOCREmbedding models
Observe · Evaluate · Govern
trace · contract-extractionarchitecture decision · why PaddleOCR + Qwen, not GPT-4Vcomplete
  1. PlanRequirement analysisExtract pricing from 200-page contracts with nested tables, merged cells, multi-currency. Commercial OCR had failed.0ms
  2. ModelEvaluated: GPT-4 VisionExcellent accuracy, prohibitive cost: $0.03/page × 200 pages × 1,000 contracts = $6,000/batch. Rejected.0ms
  3. ErrorEvaluated: Commercial OCR vendorsTested three platforms. All failed on nested tables with merged cells — critical requirement.0ms
  4. MCPSelected: PaddleOCR for table detectionOpen source, handles complex layouts reliably. Runs on existing GPU infrastructure.0ms
  5. ModelSelected: Qwen 7B for field extractionSmall enough to self-host, accurate for structured fields. Reduces cost to $0.002/page.0ms
  6. HumanDesigned: Confidence-based routingBelow 85% confidence → human review queue. Partial automation is the goal, not full automation.0ms
  7. ResultProduction outcome15× cheaper than GPT-4 approach. Client owns infrastructure. Scales linearly with volume.0ms
~/lessons-learnedproduction
# Production edge cases
1. Multi-currency tables
   → LLM hallucinated exchange rates
   → Fix: Explicit currency detection step
2. Handwritten annotations
   → OCR confidence dropped to 40%
   → Fix: Route to human, don't guess
3. Scanned at angles
   → Table detection failed
   → Fix: Deskew preprocessing pipeline
# Key insight
Production AI is 20% model, 80% edge cases.
Low confidence is a feature, not a failure.

04Approach

How I deliver.

Every phase produces a tangible artifact. No stage ends without something you can hold.

  1. 01

    Discover

    Map the decision landscape, users, and data sources.

    Problem brief + success metrics

  2. 02

    Architect

    Select patterns, define boundaries, design for failure modes.

    Reference architecture

  3. 03

    Prototype

    Ship a thin, working slice end-to-end.

    Functional prototype

  4. 04

    Evaluate

    Measure quality, safety, cost, and latency against baselines.

    Evaluation report

  5. 05

    Productionize

    Harden security, add observability, document operations.

    Runbook + guardrails

  6. 06

    Operate

    Monitor drift, run evals on cadence, iterate.

    Dashboards + review cycles

05Stack

Tools I reach for.

The technologies I use daily to build production agentic systems.

DailyProficientFamiliar
Models & inference
  • Azure OpenAI (Daily)
  • Claude / Anthropic (Daily)
  • AWS Bedrock (Proficient)
  • Open-weight (Qwen, Llama) (Proficient)
Agent orchestration
  • LangChain (Daily)
  • LangGraph (Daily)
  • MCP Protocol (Daily)
  • Copilot Studio (Proficient)
RAG & data
  • pgvector (Daily)
  • Azure AI Search (Proficient)
  • Microsoft Fabric (Daily)
  • dbt (Proficient)
Languages & frameworks
  • Python (Daily)
  • TypeScript (Daily)
  • SQL (Daily)
  • FastAPI / React (Daily)
Infrastructure & ops
  • Docker / Kubernetes (Daily)
  • Azure (Daily)
  • AWS (Proficient)
  • Terraform / IaC (Proficient)

06About

The path so far.

From full-stack development to leading AI engineering — seven years of building systems that ship.

Sathish Kumar V

I believe production AI systems should be boring — predictable, observable, and straightforward to debug at 3am. The interesting work is getting them there.

Based in Chennai, serving UK and US clients remotely. I lead a team of seven engineers, own technical direction for our AI practice, and still write code daily. Currently seeking opportunities where I can continue shipping agentic AI that runs in production — not just demos that impress in meetings.

  1. Oct 2025 – Present

    AI Solutions Architect & Team LeadNow

    JMAN Group

    Lead Cluster 5 (7 engineers) in AI Centre of Excellence. Own 15 internal AI products. Architect client solutions from pitch to production — solutions I win, I build. Five deals closed this year through technical pre-sales.

  2. Apr 2024 – Oct 2025

    AI Engineer → Product Lead

    JMAN Group

    Delivered the contract extraction project that became our flagship reference (260 hours → 26 minutes). Built enterprise AI platform now serving 200+ employees across the organization.

  3. Apr 2023 – Apr 2024

    Data Engineer

    JMAN Group

    Built medallion architecture platform processing 5M+ records daily. Implemented SCD Type 2 across 20+ entities using dbt, Redshift, and Azure Data Factory.

  4. Feb 2022 – Apr 2023

    Software Engineer

    JMAN Group

    Developed RESTful APIs using Node.js, NestJS, and FastAPI serving 1M+ daily requests. Established Docker-based CI/CD reducing release cycles by 60%.

  5. 2017 – 2022

    Full Stack Developer

    Elroi, Qserve, RmKV

    Built enterprise systems using C#, .NET, and SQL Server. Delivered retail management, accounting, and inventory solutions across multiple organizations.

07Writing & talks

Thinking out loud.

Technical writing, open source contributions, and knowledge sharing.

He transformed a vague 'we need AI' brief into architecture our engineers could build and our risk team could approve. The contract extraction system he delivered became our flagship client reference.

Delivery DirectorJMAN Group · Enterprise AI PracticeInternal performance review

08 — Contact

Let's build something agentic.

Open to conversations about AI architecture, MCP protocol, and taking autonomous agents from prototype to production. Seeking remote roles with UK/US companies — 60-90 day notice period, open to visa sponsorship.