Forward Deployed
AI Engineer
Taking private LLMs, autonomous agent swarms, and sovereign AI infrastructure from high-level business problems to production-grade deployments inside client VPCs and air-gapped clusters.
11 Repos
Full Systems
100% Private
Air-Gap & VPC
vLLM + K8s
GPU Clusters
60+ Notes
Production Specs
Kiran Machha
Forward Deployed AI Engineer
Technical Lead · Singapore 🇸🇬 · Sovereign AI Systems
The Complete Enterprise AI Production Pipeline
Enterprise End-to-End Inference & Training Flow
Input Security & Guardrails
Pre-flight safety, PII redaction, prompt injection defense, and authentication before any prompt touches models.
Core Architectural Pillars
Production Operational Guardrails
- Mask all NRIC, SSN, credit cards, and API keys before payload serialization into model context.
- Reject prompt injection embeddings with cosine similarity > 0.88 against known adversarial vectors.
- Enforce per-tenant token-bucket rate limits and role-based access control (RBAC).
Implemented Tech & Libraries
async def validate_and_sanitize_input(request: UserPrompt) -> SanitizedPrompt:
# 1. Zero-trust PII & Secrets masking
anonymized_text = presidio_engine.mask_pii(
request.prompt,
entities=["CREDIT_CARD", "PHONE_NUMBER", "GOV_ID", "SECRET_KEY"]
)
# 2. Embedding-based adversarial jailbreak check (<10ms)
is_safe = await nemo_guard.verify_prompt_safety(anonymized_text)
if not is_safe:
raise SecurityException("Prompt rejected: Adversarial pattern detected")
return SanitizedPrompt(content=anonymized_text, tenant_id=request.tenant_id)Small LLMs (SLMs) Fast Path
Fine-tuned Llama 3.2 3B, Qwen 2.5 7B, & Phi-4 for 80% of routine workflows (classification, extraction, structured JSON).
Frontier Deep Reasoning
Private DeepSeek-R1 & Llama 3.3 70B reserved for high-stakes audits, multi-step planning, and complex edge-case reconciliation.
Continuous LoRA Pipeline
Nightly Unsloth QLoRA distillation on high-confidence production traces, with automated benchmark gates before rollout.
11 Featured Production Systems
Click any project card to inspect its full architecture flow, problem/solution breakdown, API specs, and tech stack.
OpenForge AI
Self-hosted Sovereign AI platform with 24 bundled services, OpenClaw & Hermes agents, and hardware-accelerated local inference
A comprehensive, self-hosted Sovereign AI deployment platform built with 24 containerized service manifests. Bundles llama-server and vLLM local inference, OpenClaw and Hermes autonomous agent frameworks, LiteLLM gateway, Qdrant hybrid RAG, SearXNG private search, local voice STT/TTS, Privacy-Shield PII masking, and real-time GPU VRAM telemetry.
Personal AI Agent (KaviAgent)
Local-first personal assistant with SQLite memory state, local web cockpit, and 95-line plain Python loop
A local-first personal AI assistant framework demonstrating the four foundational pillars of production agents: Harness, Reasoning Loop, Stateful Memory (SQLite), and LLM-as-Judge Evals. Features a local web dashboard cockpit at localhost:7777 and Telegram bot integration.
Agentic Workspace IDE (KaviSpace)
Desktop application with Tauri v2, xterm.js terminals, and KaviSwarm multi-agent pipeline
A Tauri v2 + Rust desktop application and multi-agent development environment that autonomously writes, builds, deploys, and live-demos full-stack applications through a three-phase AI swarm pipeline (KaviSwarm).
Business Growth Agents Platform (KaviAI)
AI agent platform planning, generating, and publishing content across 10+ platforms automatically
An enterprise AI agent platform for growth, marketing, and distribution. Uses Google Gemini with LiteLLM gateway and OpenAI fallback to automate SEO, Reddit, LinkedIn, X, Instagram, and YouTube publishing.
Agentic Coding Platform & Sandboxes
Self-hosted AI development infrastructure for secure, governed agentic coding
A self-hosted developer platform providing containerized workspace environments, AI coding agents, and governance controls. Allows developers and AI agents to code side-by-side inside controlled sandbox environments.
Corporate Enterprise RAG
Production-grade Retrieval-Augmented Generation for enterprise knowledge management
A complete corporate organization RAG system that ingests internal documents, enables hybrid search across organizational knowledge, and provides intelligent Q&A through agentic retrieval with LangGraph.
Bank Cheque OCR Automation
Handwriting-aware cheque processing with automatic fraud detection for regulated banking
A production-ready Bank Cheque OCR API automating the extraction and verification of critical information from scanned bank cheques. Designed for high-volume banking back-offices processing 5,000+ cheques daily.
Customer Call Support Voice Agent
Production-ready realtime voice agent for customer service call centers
A complete customer service voice agent system capable of handling inbound and outbound telephone calls with sub-second speech recognition, intelligent tool-calling responses, knowledge lookup, and conversation tracing.
n8n AI Agents & Workflow Automation Collection
Production-ready n8n automation workflows powered by AI agents for enterprise operations
A curated collection of 17 enterprise-grade n8n automation workflows integrating LLM agents, automated triage, email notification generators, IT ticket processors, and daily reporting systems.
Invoice Automation OCR
High-throughput document parsing with vLLM, Rust API gateway, and async GPU queues
A multi-stage asynchronous invoice processing system engineered with vLLM vision model inference, a high-concurrency Rust API gateway, and async task queues for enterprise accounting teams.
PageBolt MCP Server for AI Coding Assistants
Model Context Protocol (MCP) server giving AI agents web capture, screenshots, and page inspection
An open-source Model Context Protocol (MCP) server connecting AI coding assistants (Cursor, Windsurf, Claude Desktop, Cline) to PageBolt capture APIs for screenshotting, PDF generation, and page inspection.
Kiran Machha
Forward Deployed AI Engineer
Technical Lead Engineer — Enterprise Applications
Singapore · Available Globally
Embedding inside customer engineering teams to deliver Sovereign AI & Autonomous Agents that run securely inside enterprise walls.
As a Forward Deployed AI Engineer & Technical Lead, I sit directly inside customer codebases, infrastructure clusters, and strict compliance environments to design, build, and deploy production-grade AI systems.
My core specialization is Sovereign AI & Private Infrastructure — serving and fine-tuning open-weight models (DeepSeek, Llama 3, Qwen 2.5, Mistral) inside private VPCs, air-gapped clusters, or local GPU nodes for banks, fintech, and regulated organizations.
Backed by over a decade of regulated enterprise engineering experience (banking change control, distributed microservices, zero-trust security), I bridge the gap between high-level AI ambitions and hardened, containerized vLLM serving, hybrid retrieval pipelines, multi-agent orchestration (LangGraph, CrewAI, MCP), and Kubernetes GPU autoscaling.
Business Problem → Production AI
An FDE who owns the full lifecycle from initial discovery to long-term GPU cost optimization.
Business Problem
Identify workflow bottlenecks, regulatory constraints, ROI targets, and latency requirements.
AI Architecture
Design model strategy, open vs closed weights, vector storage, context windows, and safety barriers.
Development
Fine-tune open-weight models (LoRA/QLoRA), build RAG pipelines, typed agent tools, and evaluation harnesses.
Enterprise Integration
Connect models to ERPs, core banking APIs, CRM database queues, and OAuth/RBAC identity systems.
Deployment
Containerize with vLLM/Triton, deploy on Kubernetes, GPU autoscaling, and zero-downtime rollouts.
Observability
Self-hosted tracing, evals, hallucination monitoring, semantic logging, and token usage analytics.
Production Operations
Automated failover, load balancing, continuous eval gates, and human-in-the-loop fallback queues.
Cost Optimization
Semantic caching, model distillation, dynamic batching, and GPU node pool scale-to-zero.
Integrating AI Into Operational Systems
Models create value when integrated cleanly into operational workflows, databases, and core APIs.
RAG
ProductionVector Search
ProductionOCR
ProductionNLP
ProductionVision
ProductionVoice
ProductionAI Agents
ProductionPrediction
ProductionForecasting
ProductionAutomation
ProductionOpen-Source & Self-Hosted Stack
Proven technology choices enabling enterprises to run AI models on-premises without third-party API dependencies.
AI/ML
11 ToolsLLM
10 ToolsAgents
6 ToolsInfrastructure
7 ToolsData
5 ToolsCloud
7 ToolsObservability
8 ToolsEnterprise Delivery Experience
From regulated banking change control to SaaS GPU platforms and forward-deployed client engagements.
Forward Deployed AI Engineer
Private AI & Agent Systems
- Builds and deploys agentic AI systems with customers — multi-agent orchestration, tool calling, and self-correcting reasoning loops that survive real inputs.
- Runs the full LLMOps chain: serving, gateway routing, evaluation harnesses and observability, so quality regressions are caught before customers find them.
- Ships production infrastructure on Kubernetes with GPU autoscaling, and publishes the patterns as open workshops.
AI Platform / LLM Ops Engineer
Enterprise SaaS
- Stood up the shared inference platform — model gateway, quota and usage-based billing, and semantic caching to hold cost per request down as traffic grew.
- Introduced evaluation gates into CI so prompt and model changes shipped on evidence rather than vibes.
- Moved batch and streaming inference onto autoscaling GPU node pools with scale-to-zero for off-peak hours.
Full Stack AI Engineer
Payments
- Delivered customer-facing AI features end to end — Next.js and FastAPI through to the retrieval and ranking layers behind them.
- Built hybrid search and RAG over transaction and policy corpora, with citation enforcement for audit trails.
- Hardened the path to production: async workers, idempotent retries and structured tracing across services.
Technical Lead Engineer — Enterprise Applications
Banking
- Built and maintained enterprise-grade applications under regulated change control, where a failed release is an incident report.
- Microservices and API platforms on Spring Boot and Node.js, containerised and delivered through automated pipelines.
- The grounding that makes the AI work deployable: security review, access control and data handling as defaults rather than afterthoughts.
Let's Connect
Available for Forward Deployed AI Engineer and Sovereign AI Infrastructure roles across Singapore, APAC, and EMEA.