Himansh Raj — Portfolio

Himansh Raj

AI / ML researcher building human-language technology — neural machine translation, LLM training, and vision-language models — with a focus on low-resource NLP, model evaluation, and failure-mode analysis.

Email↗GitHub↗LinkedIn↗Hugging Face↗Résumé (PDF)↗
0.00
sacreBLEU · Prakrit→English
0.0M+
Parallel pairs generated
0M+
Requests served · codegen
0.0%
Agent task pass rate
01Research

Manuscript in preparation · with IIT Delhi

Low-Resource Neural Machine Translation

English ↔ Prakrit · Prakrit → Sanskrit

Trained small English↔Prakrit translation models, then self-generated 2M+ pairs and filtered to a 1,513,093-pair corpus via word-level Jaccard stability ≥ 0.40 (beam-candidate agreement), serving inference at 498 sentences/sec with CTranslate2. Also built a Prakrit→Sanskrit lemmatizer reaching 51% character similarity from the entire available training set of just 1,696 pairs.

BLEU (sacreBLEU)47.08METEOR0.558Corpus1.51MSpeed498/s
BLEU → training47.08
aligned embedding space · en ⇄ prakrit

Ongoing · Jan 2025 – Present

Infinite-Context LLM

Active / Passive Memory Architecture

Designing an LLM/SLM memory architecture with separate active (in-context) and passive (retrievable) memory states — for scalable, near-infinite knowledge retention and efficient contextual recall without quadratic attention cost.

Memoryactive+passiveTarget∞ context
ACTIVEin-contextPASSIVEretrievable
memory statesread / write
02Selected Work & Models
03More Projects
04Experience

IIT Delhi — MISN Lab

May 2026 – Present

Research Intern · Machine Intelligence, Signals & Networks Lab

  • Building LLM-driven multi-agent recommendation pipelines (retrieval, planning, tool use) under Dr. Sandeep Kumar.
  • Developing evaluation methodology for failure-mode characterization and robustness analysis.

Colate

Jul 2025 – Dec 2025

AI / Software Engineer (Remote, Part-time)

  • Built a Vector Database + semantic search engine for sub-second retrieval over millions of code embeddings.
  • Cut inference latency 35% with profiling + Redis-backed caching; ran NLP evaluation suites for robustness.
05Stack
ML & Deep Learning
PyTorch · Transformers · Hugging Face · Fine-tuning (Llama/Mistral) · LoRA/QLoRA · RLHF · Quantization · CTranslate2
NLP & Language
Neural Machine Translation · NLU · Low-resource & Multilingual NLP · Lemmatization · RAG · BLEU/METEOR
Computer Vision
Vision Transformers (ViT) · Multimodal Alignment · Vision-Language Models · Image Feature Extraction
LLMs & Agents
Training/inference optimization · Multi-agent Orchestration · LangChain · LangGraph · MCP
Systems
Vector DBs (FAISS/Pinecone/Chroma) · FastAPI · Docker · Kubernetes · AWS · Redis · PostgreSQL
Languages
Python · C++ · C · TypeScript/JavaScript · SQL · Bash
06Education & Recognition

B.Tech, Computer Science & Engineering (AI & ML) — IILM University, Greater Noida · 2023 – Present

  • —Winner — HIVE Hackathon, IIT Delhi (2024)
  • —Winner — FinAThon, IIT Delhi (2023)
  • —Top 10 (Intl) — IIT Bombay TechFest (2023)
  • —1st — IILM Competitive Programming Challenge (2023)
  • —200+ DSA problems (Striver A2Z) on a self-built local judge
  • —Google Cloud Generative AI Apps — 6 certifications (2025)
Certifications
  • Google Cloud Generative AI Apps — 6 certificates (2025)
  • HackerRank — Python · Node.js · React · REST API
  • DataFlair — Java · Infosys Springboard — Python, C
Languages

Hindi (native) · English (professional working)

Let's build language technology.

Open to applied-science & ML research roles, collaborations, and low-resource NLP work.

iamthehimansh@gmail.comGitHub

Built by Himansh Raj · Noida, India