📍 Pune, Maharashtra, India 📧 ankitdas2k@gmail.com 🐙 github.com/nkitan 💼 linkedin.com/in/ankitdas2k

I'm a Systems Software Engineer and Site Reliability Engineer with 4+ years of experience building and operating highly available distributed systems.

At PhonePe, I work on infrastructure responsible for handling 1.4M+ QPS and supporting systems serving roughly 60% of India's UPI transaction volume. My work spans edge traffic infrastructure, distributed configuration management, CDN/DNS automation, workflow engines, observability, incident response, and AI-assisted operations.

My engineering interests sit at the intersection of:

• Distributed systems and infrastructure
• Systems programming
• Site reliability engineering
• AI agents and autonomous systems
• Local-first AI
• Developer and operations tooling
• High-performance backend systems

I primarily work with Python, Rust, C/C++, TypeScript, Linux, PostgreSQL, NGINX, Prometheus, and modern LLM infrastructure.

I also build end-to-end products outside work—from browser-native AI inference and autonomous-agent runtimes to model-training systems, video-production software, homelab infrastructure, and business platforms.

February 2023 – Present · Pune, India

I work on production infrastructure responsible for some of India's highest-volume financial traffic, combining software engineering with reliability, networking, distributed systems, and operational automation.

Architected a distributed configuration-observability platform covering 2,000+ NGINX servers.

The platform asynchronously collects proxy configurations and host telemetry using FastAPI, RabbitMQ, and Celery, and provides:

• SHA-256 deduplicated configuration history
• Intra-region and cross-region drift detection
• Fleet-wide configuration search
• Git-backed configuration discovery
• Historical configuration inspection
• Secure multi-file comparison
• Shareable configuration diffs
• Slack-based drift alerting

Configuration history is indexed in Elasticsearch, with a frontend built using React, Tailwind CSS, and shadcn/ui.

Architected and developed a provider-agnostic SRE control plane for safely managing CDN and DNS infrastructure.

The system provides:

• Domain validation
• Approval-gated infrastructure changes
• Durable execution state
• Phased deployments
• Provider audit trails
• Failure recovery
• Disaster-recovery configuration
• Extensible provider contracts

Implemented Cloudflare as the reference provider, including automation for:

• Zones
• DNS records
• Origin pools
• Load balancers
• Disaster recovery

The architecture enables parallel integrations for Akamai, AWS Route 53, and other providers without coupling the core platform to individual vendors.

Designed and built a visual workflow platform that lets engineering teams create operational automation without modifying the core application.

The platform includes:

• 19 composable node types
• Versioned workflow DAGs
• Manual triggers
• API triggers
• Scheduled execution
• Conditional routing
• Parallel execution
• Human approval gates
• Reusable UI panels
• Workflow result sharing
• Vault-backed secret management

The goal is to turn repetitive operational procedures into reusable, governed software.

Built a horizontally scalable workflow execution engine backed entirely by PostgreSQL.

The runtime supports:

• Multi-process workers
• Durable queues
• Immutable workflow versions
• Pause/resume
• Crash recovery
• Concurrent execution
• Persistent execution state

The architecture removes the requirement for a separate Redis or Celery broker while retaining durable asynchronous execution.

Built an internal on-call copilot for infrastructure investigation.

The system asynchronously ingests engineering knowledge from Confluence and Jira and exposes operational tools capable of querying:

• InfluxDB telemetry
• CDN telemetry
• Proxy configuration state
• Configuration drift information

I also exposed internal infrastructure capabilities through MCP tools and created reusable AI skills capable of generating operational workflows.

The objective is not merely conversational AI—it is giving models safe, structured access to real operational systems.

Alongside software development, I own production responsibilities around edge traffic infrastructure, including:

• Incident response
• SLA/SLO-driven reliability
• Traffic management
• Performance optimization
• Disaster recovery
• Configuration management
• Observability
• Capacity and failure analysis
• Operational automation

I am particularly interested in building AI systems that can actually perform useful work rather than limiting LLMs to conversational interfaces.

My work focuses on:

• Autonomous agent runtimes
• Multi-agent orchestration
• Human-in-the-loop systems
• Tool calling
• Model Context Protocol
• Retrieval-Augmented Generation
• Evidence-grounded generation
• Long-term agent memory
• Model routing
• Local inference
• Browser-native inference
• Workflow generation
• Agent sandboxing
• Approval-gated actions
• Model training and optimization

I frequently experiment with both hosted and local models and build systems around OpenAI-compatible APIs, llama.cpp, Ollama, LangGraph, FastMCP, WebGPU, and WASM.

A local-first autonomous-agent runtime designed to turn incoming requests into governed, executable work.

Requests can arrive through:

• Chat
• API
• CLI
• Web
• Desktop

The runtime can select an appropriate model, recall relevant long-term memory, execute LangGraph tool loops, and decompose complex tasks into parallel subagents.

Safety-critical operations are approval-gated, while generated code executes inside restricted containers.

Its autopilot can also ask a model to generate a bounded workflow, validate that workflow, repair invalid plans, and execute the resulting graph.

A local-first platform for managing fleets of AI agents through a tmux-style terminal interface and web UI.

Features include:

• Remote agent management
• Workflow scheduling
• Token-aware model routing
• Usage tracking
• Multi-agent coordination
• Centralized execution visibility

A local-first, multi-tenant system that converts an objective and supporting material into governed deliverables.

The platform can:

1. Plan research
2. Collect web and uploaded evidence
3. Delegate tasks to specialized agents
4. Fact-check intermediate work
5. Synthesize final outputs
6. Generate landing pages or images
7. Stream the entire execution process to a dashboard

Its runtime incorporates:

• Tenant isolation
• Budgets
• Audit events
• Approval gates
• Evidence provenance

Governance is implemented in the execution system itself rather than relying solely on prompts.

A local-first, evidence-bound AI video production studio.

The platform manages the complete production lifecycle:

• Brief confirmation
• Research
• Evidence-backed scripts
• Storyboards
• Generated/imported media
• Asset review
• Audio layers
• Frame-accurate editing
• FFmpeg rendering
• Publication packages
• NLE interchange
• Portable .laivga project archives

SQLite acts as the authoritative project state, while assets are content-addressed to maintain provenance.

A core design principle is deterministic behavior: the application does not silently switch models or workflows when something fails.

A sequence of experiments in building and validating local language-model training systems.

The mature training pipeline supports:

• Small-model training
• Exact evaluation verifiers
• Bounded improvement cycles
• Sentinel rollback
• Live browser training visualization
• Deterministic dataset lineage
• Data prefetching
• Multiple optimizer strategies
• Crash recovery
• Distributed training
• Hardware telemetry
• Quantization-aware training
• Pruning
• Knowledge distillation
• Inference optimization

Performance experiments have been conducted on an RTX 4070 Laptop GPU, with measurements reported alongside workload details.

A local AI model manager focused on organizing and running models from the same model family.

Features include:

• Structured model storage
• Model downloading
• llama.cpp integration
• llama.cpp TurboQuant support
• Runtime parameter optimization
• Integrated AI agent functionality

A reproducible infrastructure platform designed to turn a low-power ARM single-board computer into a capable homelab.

A modular installer handles:

• Host hardening
• Networking
• Storage
• Podman installation
• Rootful LAN infrastructure
• Rootless application services
• Boot automation
• Operations tooling

A compact Rust API and web interface expose controls for:

• Containers
• Pods
• Files
• Schedulers
• Knowledge bases
• Calendars
• Kanban boards
• System administration
• AI agents

A multi-tenant, AI-native business operating platform in the Salesforce / Zoho / Odoo category.

It combines:

• CRM
• Work management
• Inventory
• Invoicing
• HR and payroll
• Analytics
• Marketing automation
• Social automation
• Company knowledge
• Cross-domain workflows

Its AI agent executes operations as the current user, meaning every action passes through normal RBAC controls.

Sensitive operations support:

• Preview-before-execution
• Human approvals
• PII redaction
• Tenant-level AI kill switches
• Platform-wide AI kill switches

Browser-native voice generation using WebGPU and WASM, allowing compatible models to execute locally without requiring traditional server-side inference.

Browser-based LLM execution using WebGPU/WASM, exploring client-side inference as an alternative to centralized inference APIs.

A lightweight utility for detecting and inspecting browser WebGPU support.

A GNOME desktop utility for theme management and automatic dark/light mode scheduling.

Built around native Linux desktop technologies and designed to integrate cleanly with the GNOME environment.

January 2022 – July 2022 · Remote

Led the design and development of a distributed web scraper using:

• Java
• Hibernate
• Selenium
• MySQL
• AWS

The system ran on EC2 Spot Instances and was designed around idempotent processing so workloads could recover from instance preemption.

The resulting architecture reduced infrastructure cost while retaining fault tolerance and operational simplicity.

Indian Institute of Technology (IIT), Roorkee June 2024 – December 2024

Focused on modern Generative AI and machine-learning systems, including deep learning, LLMs, NLP, computer vision, and production AI architectures.

Symbiosis Institute of Technology, Pune June 2019 – May 2023

Relevant coursework included:

• Deep Learning
• Computer Networks
• Data Structures & Algorithms
• Computer Organization
• Compiler Construction
• Python Programming

I enjoy problems where software engineering, infrastructure, and AI intersect.

In particular:

Distributed Systems · Systems Programming · SRE · AI Agents · LLM Infrastructure · Local AI · Observability · Networking · Developer Tooling · Workflow Engines · Model Optimization · Human-in-the-Loop Automation

Build systems that are observable, recoverable, understandable, and difficult to misuse.

Reliability is more than uptime. A system should make failures visible, recover predictably, preserve enough state to explain what happened, and remain operable by humans when its abstractions inevitably break.

The same principle applies to AI systems: autonomy is useful only when paired with evidence, observability, bounded execution, and human control.