I build human-language technology — training and fine-tuning LLMs / SLMs, neural machine translation, and vision-language models — and I study model behavior, robustness, and failure modes in learning-based systems.
- 🔬 Research Intern @ IIT Delhi — MISN Lab (agentic systems for recommendations)
- 🤖 AI Engineer @ Colate — agentic IDE, multi-model codegen proxy, vector search
- 🎓 B.Tech CSE (AI & ML) @ IILM University
- 🌐 Everything lives at himansh.in — research, projects, blog, and a little 3D game
Low-Resource Neural Machine Translation — English ↔ Prakrit, Prakrit → Sanskrit Manuscript in preparation · joint work with a PhD researcher at IIT Delhi
Trained small English↔Prakrit models, then self-generated 2M+ pairs and filtered them to a 1,513,093-pair corpus via a word-level Jaccard-stability check — reaching BLEU 47.08 (sacreBLEU) / METEOR 0.558 (Prakrit→English) at 498 sentences/sec with CTranslate2. Also built a Prakrit→Sanskrit lemmatizer (51% char-similarity from just 1,696 training pairs).
↳ prakrit_eng_translator · prakrit dataset
Infinite-Context LLM — an active (in-context) / passive (retrievable) memory architecture for near-infinite recall without quadratic attention cost. (ongoing)
| Project | What it is |
|---|---|
| 🖼️ LlamaVision | 6.3M-param image projector aligning a ViT encoder to a Llama-3 text embedding space — studies cross-modal alignment. |
| 🧠 F75 | A 107K-param transformer LM trained end-to-end on a single GPU — tokenizer, training loop & inference from scratch. |
| 🤖 Cana · @ Colate | Agentic IDE + secure multi-agent runtime (OpenAI / Anthropic / Gemini / Kimi) on hardened AWS K8s — 98.2% task pass across 100+ users; a compiler-style multi-model codegen proxy has served 900M+ requests. |
| 📄 BlackTable | Production RAG document-analysis API across 4 formats (PDF / MD / Word / text via Docling). |
| ⚙️ H++ | A statically-typed language built on a single primitive: bit — all types are zero-cost bit-level abstractions. |
| 🧮 STRIVER-JUDGE | A self-built local DSA judge — 200+ Striver A2Z problems, auto-validated against test cases. |
| 🕹️ 8085 Simulator | Full Intel 8085 emulator — 64KB memory, instruction-level breakpoints, step-through. |
| 🛡️ NetGuard | No-root Android firewall — per-app + wildcard DNS blocking via a local VPN tunnel. |
| 🧠 Doremon | EEG-controlled smart home — 67% intent accuracy on 2h of hackathon data. |
🥇 Winner — HIVE Hackathon, IIT Delhi (2024) • 🥇 Winner — FinAThon, IIT Delhi (2023) • 🌍 Top 10 (Intl) — IIT Bombay TechFest (2023) • 🧠 1st — IILM Competitive Programming (2023) • ☁️ Google Cloud Generative AI (6 certs)