| Model | ish-codes-magic/ishita-pal |
|---|---|
| Role | Founding Engineer & AI Lead at Storefox.ai, leading the audio-intelligence pipeline that turns raw in-store conversations into insights for retail teams |
| Intended use | Taking messy real-world signals (noisy audio, pitch-deck PDFs, terabytes of logs) and building LLM systems around them that are evaluated, cheap to run, and trustworthy |
| Pre-training | Research at DESY, CERN and IISER Pune → face recognition at scale at Coriolis → LLM hiring platforms and real-estate chatbots at Hashkraft |
| Evals | 80% accuracy across insight-generation tasks for 15+ retail clients · audio pipeline made 2× faster and 5× cheaper ($100 → $20 per audio-hour) |
| Alignment | Whitebox Research AI Safety Fellowship: interpretability and chain-of-thought faithfulness |
| Known limitations | Reads books, listens to music, and still wonders when AI will take my job |
| Project | What it does | Built with |
|---|---|---|
| self-recognition-replication | Replicates LLM Evaluators Recognize and Favor Their Own Generations. GPT-4.1-nano picked out its own summaries 92–96% of the time and preferred them 99% of the time. | OpenAI API, logprobs |
| transformer_from_scratch | Attention Is All You Need with every component hand-built in PyTorch, trained for English → French translation | PyTorch, HF Tokenizers |
| IPO-Readiness-PDF-Analyzer | Reads SME pitch decks with a vision LLM, scores them on 8 IPO-readiness criteria and writes up the risks and gaps | Next.js, FastAPI, Gemini |
| GPT_art_critique | Upload an artwork, get back a structured critique with per-criterion scores | GPT-4 Vision, Instructor, Streamlit |
| DESY_Summer_Project | Anomaly detection for DESY's dCache storage system over ~10 TB of transfer telemetry (85% accuracy) | PySpark |
🧱 I learn by rebuilding things from scratch: LLM internals (attention, tokenisers, positional embeddings, GPT) · CV models · ML algorithms · AlexNet · VGGNet · BPE tokenizer
- Gradient Descent Algorithm: How Does it Work in Machine Learning?
- Intuition behind perceptron: the building blocks of Neural Networks
- 20 Most Asked Interview questions on Python
- Heap sort explained using python
- Webscraping using Scrapy: Creating your first scrapy project
| Languages & ML | |
| Backend & data | |
| Cloud & infra | |
| LLM tooling | Instructor · Pydantic · Portkey · Hugging Face · OpenAI · Claude · Gemini |
>>> ishita.generate("Can we work together?")
'Yes! LinkedIn or email are the fastest ways to reach me.'