making machinesthink
0%
Bangalore, Indiaopen to interesting problems

I make machines think and interfaces feel.
Currently: AI Engineer — depends on the hour.

View selected work About me

senior swe @ stride.ai

icpc regionalist / top 0.3% leetcode

scroll
01Captar
click me · next projects
LLM Orchestration✦Agents✦RAG✦LoRA Fine-Tuning✦TypeScript✦Python✦Rust✦C++✦Next.js✦Node.js✦PostgreSQL✦Docker✦AWS✦Solidity✦Eval Pipelines✦
LLM Orchestration✦Agents✦RAG✦LoRA Fine-Tuning✦TypeScript✦Python✦Rust✦C++✦Next.js✦Node.js✦PostgreSQL✦Docker✦AWS✦Solidity✦Eval Pipelines✦

about me · under the hood

One engineer.
four ways of thinking.

I move between agent infrastructure, full-stack products, model systems, and lower-level engineering. The medium changes; the habit stays the same: understand the constraint, build the smallest useful system, then measure what happens in the real world.

portfolio / about / engineering-modesReady
Agent InfrastructureTools · Guardrails · Evals
orchestrationpolicyevalsbuilding · measuring · learning
·about.devansh — type `help`, or choose a mode
❯open agents
✓agent infrastructure ready — tools, policy, traces
FIG. 001 — Four modes, one engineerAgents, products, models, systems — explore one above

Iamanengineeratheart.Iobsessovercleansystems,butIloveitevenmorewhenthecodeactuallymakespeoplefeelsomethingandgetthingsdone.

beautiful systems that actually ship.

Primarily focused on

A few years of shipping, settled into three tracks that sharpen each other every day.

20+ TOOLSAGENTS

AI Engineering

2+ years

Agents, guardrails, evals — LLM systems that behave in production instead of demos.

AI is part of how I design & build,
every day.

Not a novelty — a daily practice. These are the tools I reach for to move from a rough idea to a shipped, working product.

Claude01
Codex02
Cursor03
Vercel04
Grok05
GitHub06
Ollama07
DeepSeek08
Tailwind09
OpenRouter10

I'm an open book. here's the unfiltered timeline.

2023

Intern at Cloudbloom

Started shipping React components and squashing 30+ defects before anyone trusted me with prod. ACM ICPC Regional Finalist the same year — sleep was optional.

2024

Promoted. Graduated. Repeated.

B.Tech from IIIT Dharwad. Promoted to Software Developer after leading frontend delivery. Built an LMS serving 1,000+ concurrent users. Also taught DSA to 100+ students weekly at GeeksforGeeks.

2025

Senior Engineer at Stride.AI

Now I design LLM-agent workflows end to end — orchestration, tools, product surface. 20+ features owned, API layer scaled to 10K+ requests/day, founders on speed dial.

now

Building at midnight

Captar, Baelys, PhoneClaw, a vector DB from scratch in Elixir, a terminal browser in Go. Some people doomscroll. I ship side projects.

ship it✦break it✦fix it✦ship it again✦
ship it✦break it✦fix it✦ship it again✦

Selected Projects worth a look

01

01 — Runtime control for LLM applications

Captar↗

A guardrail layer that wraps the OpenAI client — per-session budget reservations, tool allowlists, execution policy. No proxy, keys never leave the host app. Span-first tracing with violation capture and eval-ready dataset export.

  • TypeScript
  • Next.js
  • PostgreSQL
  • OpenAI API
02

02 — Multi-model AI image editor

Baelys↗

20+ LoRA style-transfer models, AI background removal, and a canvas-based multi-layer editor. Launched on Product Hunt, monetized with tiered subscriptions, and publicly praised by Vercel's CEO.

  • Next.js
  • Node.js
  • LoRA
  • Canvas
03

03 — Phone automation, fully local

PhoneClaw↗

End-to-end Android automation inspired by OpenClaw — an agent that drives your phone, running entirely on-device. No cloud, no leaks, just a phone doing what you tell it.

  • TypeScript
  • Android
  • Agents
04

04 — A vector database from scratch

elix-db↗

Custom vector DB built in Elixir to explore indexing architectures — HNSW, IVFFlat — from first principles. Because the best way to understand a database is to build one.

  • Elixir
  • HNSW
  • IVFFlat
05

05 — A browser that lives in your terminal

Drowser↗

Terminal-based browser engineered in Go, driving Brave in headless mode. The web, rendered where developers actually live.

  • Go
  • Headless
  • CLI

more experiments rotting in public on github

AI does not ship.
It does the other 80%

Using AI across the daily workflow for faster execution, better decisions, and shipping real-world products.

01orchestration

LLM Agents & Tools

Multi-step agent workflows with tool calling, wired into real product surfaces — not notebook demos.

02guardrails

Runtime Control

Budget reservations, tool allowlists, execution policy. The keys stay inside the app, always.

03evals

Traces & Scoring

Span-first tracing, violation capture, weighted rubrics — regressions caught before users find them.

04models

LoRA & RAG

20+ style-transfer LoRAs running in production, retrieval only where it actually earns its keep.

05daily kit

Claude · Cursor · GPT

Pair programming with agents, PR review loops, MCP servers sharing context in one loop.

Let's build
something weird

Ready to make a move? Drop an email to discuss roles, collaborations, agent infrastructure — or just to say hi. Response within 24 hours.

say hi before overthinking it→
01Claudedaily kit