Skip to content

Learning in Public

Engineering Philosophy

Great systems are built through iteration, and great engineering is learned in public. I share my technical experiments, open-source contributions, and learnings along the way.


File-To-BinaryVideo-BackTo-File Golang

Engineered an encoding mechanism that converts any file into a binary video format, enabling lossless decoding back to the original source. An exploration of binary data representation, video encoding pipelines, and creative approaches to data storage.

View on GitHub →

Pub/Sub Implementation using Redis Golang

A step-by-step technical breakdown of building a highly available publish/subscribe messaging system using Redis. Covers connection management, message serialization, and patterns for reliable message delivery.

Read the Write-up →

AI-Assisted Engineering Practice AI / Agentic Workflow

Through building PhotonicOps, I developed hands-on experience running an AI-augmented engineering workflow at a level beyond simple autocomplete:

  • Agentic project context: Configured project-level constraint files (CLAUDE.md, .agents/AGENTS.md) to automatically enforce architectural rules (ARM64-only Docker configs, zero-allocation Go patterns, zero-cloud-API policy) across every Claude Code session — no manual reminding needed.
  • Directory-scoped agent personas: Designed custom AI skills (go-architect, dsp-math, mlops-agent), each carrying domain-specific constraints (e.g., "vectorized NumPy only, no for loops over data arrays") so the agent self-enforces the right rules per part of the codebase.
  • ADRs as machine-readable context: Used Architecture Decision Records as durable context so an AI agent picking up work later doesn't reinvent or contradict prior design decisions.
  • Gate-checked phased roadmap: Structured a phased roadmap (Phase 0 → 1 → 1.5 → 2 → 3) with explicit, testable gate criteria (e.g., "zero significant GC pauses," verified via pprof) to keep AI-assisted work scoped and verifiable rather than sprawling.

AI-First Approach to Learning AI / Agentic Workflow

I've adopted an AI-first methodology for continuous learning and engineering — using agentic tools not just to write code faster, but to explore new architectures, prototype ideas at speed, and learn by doing:

  • Claude Code — agentic/autonomous coding with directory-scoped project context and constraint enforcement.
  • Google Antigravity — advanced agentic capabilities for complex multi-file refactoring and problem-solving across a full codebase.
  • GitHub Copilot for Enterprise — enterprise-context-aware inline generation for accelerated day-to-day coding.
  • OpenRouter — multi-model API access for comparing providers (cost/latency/context-window/capability tradeoffs) without vendor lock-in.
  • Ollama — self-hosted local LLM inference for offline/air-gapped environments (used in PhotonicOps).
  • Langfuse — LLM observability and tracing for agentic workflows.

What's Next

I have recently started exploring AI Platform Engineering and MLOps, learning how backend principles of reliability and scalability apply to AI infrastructure — from prompt pipelines and model serving to LLM observability. Early days, but sharing the journey as I go.