02 · About — The Drafting Workbench
The Drafting Workbench
The one sunlit room in the cabinet: paper, rules and pins, where the machines are sketched before they are built.
Yash Mishra
AI/ML Engineer & Researcher
Delhi NCR
Underneath the abstraction
I like understanding what happens underneath the abstraction. Whether I'm implementing a research paper from scratch, wiring a retrieval system, or testing how an AI agent fails, I want to know why it works — and, more usefully, where it breaks. That habit is why most of the machines in this cabinet exist: not to wrap a library, but to open it up.
The work I reach for sits where AI/ML meets real evaluation — retrieval, agents, LLM security, efficient Transformers. Building the system is half of it; measuring what it actually does is the other half. If a claim can't be tested against something real, I'd rather not make it.
Everything here was made the same way: sketched first, implemented honestly, evaluated against evidence, and shipped with its limits labelled.
Education — the foundation
BTech — Computer Science and Engineering (AI & ML)
Greater Noida Institute of Technology
2022–2026 · graduated June 2026
The GATE arc — persistence, on the record
GATE 2026 — qualified, both papers
Computer Science & Data Science — both qualifiedAIR ≈ 6,000 (approximate All India Rank, resume-sourced)
GATE 2027 — preparation is on.
The tool wall
Grouped by what they do, hung where I can reach them. No percentages, no bars — a tool earns its hook by being used.
Languages
- Python
- C / C++
- JavaScript
- TypeScript
- SQL
AI/ML & NLP
- PyTorch
- TensorFlow
- Transformers
- Hugging Face
- scikit-learn
- SentenceTransformers
- Model training & fine-tuning
- Neural Machine Translation
- OpenCV
GenAI & Retrieval
- RAG
- Chroma
- Qdrant
- Embeddings
- Gemini
- OpenRouter
- LangChain
- OpenAI
AI Tools & IDEs
- Codex
- Claude Code
- Claude
- ChatGPT
- Antigravity
- Cursor
- VSCode
- OpenCode
- Ollama
- Higgsfield
- ElevenLabs
- Perplexity
AI Evaluation & Security
- Agent evaluation
- Adversarial testing
- Prompt-injection defence
- Retrieval evaluation
- BLEU / chrF / METEOR
- Interpretability
- LLM security & access control
Backend & Web
- FastAPI
- Next.js
- React
- REST APIs
- PostgreSQL
- SQLite
- Prisma
- SQLAlchemy
- Authentication
Cloud & Deployment
- AWS
- Google Cloud Console
- Cloudflare
- Netlify
- Firebase
- Vercel
- Render
- Supabase
- Railway
Tools & Testing
- Git / GitHub
- Docker
- Postman
- npm
- pytest
- Vitest
- Playwright
- Ruff
- Jupyter
- NumPy
- Pandas
- Matplotlib
From the profile
- Stable Diffusion
- Gradio
Hanging where I can reach them. The wall says what I use — not how well, and not since when: proficiency is what the shipped machines in Projects demonstrate, not a label pinned here.
The research, pinned to the bench
The published chapter, and the schematic that explains how the machine reads one language and writes another.
Research — the Transparent Transformer
sheet 12-A · conceptual schematic · not to scaleencoder stack × N
the source, understood
decoder stack × N
the target, written one token at a time
the encoder's output becomes the keys and values of the decoder's cross-attention — the bridge between reading and writing
The schematic is the conceptual architecture we worked from — the published chapter documents the exact implementation. No layer counts, dimensions or scores are invented here.
My published chapter builds an English–Hindi neural machine-translation system around a custom Transformer — implemented from scratch and evaluated on the IIT Bombay corpus (cfilt/IITB). Positional information comes from sinusoidal encodings; attention does the heavy lifting, and the framework was designed so the evaluation could actually see inside it.
A Transparent Transformer Framework for English–Hindi Neural Machine Translation: Design and Preliminary Evaluation — ICTCS 2025 · Springer Nature · Lecture Notes in Networks and Systems, vol. 1877. The journal version is in preparation — not published, and not presented as if it were.
the published chapter's full recordThe personality shelf
Small things, honestly held — the bench is not only for work.
Cricket
Virat and Rohit — the highlights reel runs on loop.
Pool
been playing it for a while now — and I absolutely love it.
Chess
grinding so hard, just to learn there are greater overthinkers than me.
Music
Spotify runs while everything else compiles.
History
ancient and modern-day — exploring both ends of the timeline.
Food
how could I leave it out? Planning the next meal is a legitimate discipline.
That is the bench. The machines it produced hang in Projects, the dated record sits in Experience, and if any of it raises a question, the letter slot is always open.