TRAKSHAN MISHRA · FULL-STACK & AI/ML

Wander through my work.

WASD to walk · hold SHIFT to sprint · joystick on touch · visit all five gates
STATION
—
—
QUEST
Visit the five gates
SCROLL FOR THE CLASSIC SITE ↓
TOKEN REDUCTION
10,747-69.5%
BENCHMARK PRECISION
79.6%
OCR ACCURACY
94%
01

Selected work

Hover a card — the background comes alive. Click DiffContext to play the real thing.

FEATURED · 2026 · CONTEXT ENGINE · ▶ CLICK TO PLAY

DiffContext

A repository-scale context-selection engine for code LLMs. A sys.settrace() tracer builds gold-standard execution graphs, then prunes prompts from 35,270 → 10,747 tokens while preserving task-relevant context.

PythonASTDependency GraphsLLM EvaluationAgentic Coding
GitHub ↗
-69.5% tokens79.6% precision
requests bench · 249 nodes · 181 edges
drag me
FINTECH · 2025

TradeTrack Pro

94% OCR

Real-time market data, localized GST tax calc, OCR bill-scanning & a normalized Postgres schema behind OAuth.

ReactFastAPIPostgreSQLCoinGeckoOCROAuth
STREAMING · 2025

Movie / TV Platform

mobile-first

Full-stack streaming platform with TMDB integration, auth, and a content-delivery / ad-management backend.

React.jsNode.jsTMDB APIPostgreSQL
HOW DIFFCONTEXT THINKS · KEEP SCROLLING

An agent's context window is a landfill.

Trace what the interpreter actually runs.

Delete everything else.

35,270 prompt tokens −69.5%
LIVE · VERIDIA veridia-bpq.pages.dev

Veridia — explore the world.

A playable open-world demo built entirely in vanilla JS + WebGL — no engine, no framework. WASD to walk, visit all five gates to read the story, and rest by the tavern. Deployed on Cloudflare Pages.

WebGLVanilla JS3D AssetsCloudflare Pages
02

Experience

Software Engineer

NextGen Invent Sep 2024 — Mar 2025
  • ›Shipped production full-stack apps (React.js, Django REST, FastAPI); improved API response times 35%.
  • ›Built accessible Tailwind UIs with OAuth-secured APIs across concurrent client projects.
  • ›Owned features end-to-end — schema design, REST endpoints, UI — across multiple client codebases in parallel.

AI Engineer Intern

Modern Monk Labs Jun 2024 — Sep 2024
  • ›Built FastAPI inference endpoints handling 150 concurrent req/s.
  • ›Integrated PostgreSQL pipelines for AI-output storage/monitoring; optimized model-serving latency.
  • ›Load-tested and profiled serving paths to keep p95 latency stable under concurrent inference traffic.

Generative AI Engineer Intern

VRARMR Apr 2024 — Jun 2024
  • ›Built NL2SQL models; A/B tests improved multilingual transcription accuracy to 76%.
  • ›Automated workflows with LangChain agentic pipelines.
  • ›Ran structured A/B evaluations across model variants to pick production candidates on accuracy-per-cost.
03

Toolkit

LANGUAGES & FRAMEWORKS
PythonTypeScriptJavaScriptReact.jsNext.jsNode.jsFastAPIDjangoDjango RESTTailwind
AI · ML · GENAI
LangChainLangGraphAgentic AIRAGHF TransformersOpenAI APIsLLM EvalTensorFlowPyTorchNLPComputer VisionSAP AI Core
DATA & ANALYTICS
PostgreSQLSQLData PipelinesData AnalysisREST APIsReal-Time Market APIsOCR
INFRASTRUCTURE & TOOLS
DockerGitOAuthCI/CDCloud Deployment
04

About

I care about resource-efficient AI — production systems that do more with less: AST-based context selection, token pruning, and agentic pipelines that hold up under real budget and latency constraints.

I ship full-stack web apps with React, Django REST and FastAPI, and build Generative / Agentic AI systems with LangChain, LangGraph and RAG. Recent work includes SAP AI Core feedback workflows benchmarking Gemini, Mistral & GPT in production, plus LangChain RAG-retrieval contributions and Gradio demos for text-to-image models.

05

Let's talk

Let's build something efficient.

→ trakshanmishra477@gmail.com → LinkedIn → GitHub
opens your email app with the message pre-filled
try: whoami · contact · sudo hire
Last login — portfolio session. Type a command.
trakshan@portfolio:~$
© 2026 Trakshan Mishra · Noida, IN Built with restraint. More signal, fewer tokens.