Hi, 👋
I'm Sujit.
My work sits across Machine Learning, Deep Learning, NLP , GenAI , LLMs, agentic systems, retrieval, evaluation, computer vision, and product engineering. I care about what happens after the demo: messy inputs, imperfect models, latency, observability, cost, reliability, and the engineer who eventually has to operate the thing. and love to make more progress in GenAI
I've worked on systems that moved revenue, user engagement, and engineering velocity — not just benchmark scores.
Portfolio · Writing · LinkedIn · X
| Result | Where | What |
|---|---|---|
| 1st 🥇 | WhipScribe Buildathon 2026 | CallCoach-AI: four agents score a call, and every flag plays from the second it was said. |
| 2nd 🥈 | Datathon 2025, 2nd runner-up | Designed scalable computer vision system having ML backend for retail stores like dmart and wallmart using computer vision for item price , monitoring and to improve sales of the retail shop using computer vision showing high scale AI implementations taking vision model accuracy from 45% → 79% accuracy. |
| Top 30 nationwide | Hackspiration '24 | HarvestConnect, a farm-to-buyer marketplace. |
| Shortlisted for the finals | 100x Engineers GenAI Hackathon 2024 | Dataviz AI turns raw data into animated infographic videos. |
| Shortlisted, internal round | Smart India Hackathon 2024 | ArogyaKrishi, crop disease and soil analysis. |
| 300+ users | CallCoach-AI | Founders and sales teams scoring their calls. Real traffic: enough of it that we ran through our API quota. |
| +2% revenue | Launchpad.AI | Multi-agent and Graph-RAG systems shipped to a live customer. I ran the retrieval evals and the A/B ranking tests. |
| 30 PRs merged | Open source, Oct 2026 | pandas-stubs, typeshed, MLflow, DSPy, ONNX, Hugging Face Hub and Kestra. |
| 30+ contributors | HeadOn | Coordinated the training-data labelling and ran the precision, recall and F1 evals that decided what to fix next. |
CallCoach-AI scores a founder's investor call with four agents. Every flag has a button that plays the recording from the second it was said, so you can check it yourself. It won the WhipScribe Buildathon in October 2026, first of 52 builders. Code · How it works
ArogyaKrishi ArogyaKrishi is my most-starred repository, built with Team Anant for Smart India Hackathon 2024 (SIH1638). An AI-powered farming assistant that helps farmers detect crop diseases, understand soil health, and get personalized crop and fertilizer recommendations through simple image uploads. We onboarded local farmers to test our MVP in real-world conditions, turning their feedback into actionable improvements. Our mission: make AI-driven agriculture accessible, empower farmers with smarter decisions, and promote sustainable farming.
Sentinel-AI sits next to an LLM and scores every prompt and response from 0 to 100, with the claims and flags behind the score, then decides whether to serve, fix or block the answer. The Python SDK is on PyPI as sentinelai-risk. Code · How it works
Most of these started the same way: I ran a checker against real code and fixed what it found. For pandas-stubs that meant type-checking 1,116 runnable pandas docstring examples with mypy, pyright, pyrefly and ty. For typeshed it meant running stubtest against the real packages.
- My LLM judge quoted the call. I couldn't find a third of its quotes. (Oct 2026)
- From hospital bed to hackathon: Datathon 2025 (Mar 2025)
- How I started contributing to open-source ML (Oct 2024)
More at sujit.top/notes.
What's actually going on inside these models, plus two physics problems I can't leave alone.
| Area | The question | Papers |
|---|---|---|
| Consciousness and introspection | Do models know anything about their own internal states? | Emergent introspective awareness (2026) Mechanistic study of introspection (2026) Introspection is content-agnostic (2026) Subjective experience under self-reference (2025) LLMs are aware of their learned behaviors (2025) Disproof of LLM consciousness (2025) |
| Hallucination | Why do models make things up, and can we see it coming? | Why language models hallucinate (2025) H-Neurons (2025) Deception vs hallucination (2026) Topological signatures of hallucination (2026) Uncertainty vs hallucination (2026) |
| RL and environments | What do models learn when trained with RL inside environments? | DeepSeek-R1 (2025) Does RL add reasoning beyond the base model? (2025) Absolute Zero (2025) Agent World Model (2026) RL world model learning (2026) |
| Human-like behaviour | How close is model behaviour to human cognition? | LLMs pass the Turing test (2025) Emergent misalignment (2025) CogToM (2026) Human-like reasoning in LLMs (2026) Reasoning aligns LMs to human cognition (2026) Understanding LLMs (2026) |
| LLMs from scratch | Can I build and run a coding LLM end to end on my own machine? | Attention Is All You Need (2017) FlashAttention (2022) GPTQ (2022) AWQ (2023) QLoRA (2023) vLLM (2023) |
| Black holes | What happens where general relativity and quantum mechanics don't fit? | Hawking radiation (1975) Page curve (1993) Firewalls (2012) Islands (2019) Entropy of Hawking radiation (2020) EHT M87 image (2019) |
| Dark matter | Overwhelming indirect evidence, still no direct detection. Why? | Rubin and Ford (1970) Bullet Cluster (2006) Planck 2018 (2018) LZ search (2024) |
If my commit graph goes quiet for a week, I'm probably stuck inside a Penrose diagram.
sujit/
├── models/ PyTorch · TensorFlow · Hugging Face Transformers · scikit-learn · OpenCV
├── llm-systems/ LangGraph · CrewAI · LlamaIndex · Model Context Protocol · Ollama
│ ├── retrieval/ Graph-RAG on Neo4j · FAISS
│ └── evals/ LLM-as-judge · verbatim-quote checks · precision / recall / F1 · A/B tests
├── backend/ Python · FastAPI · PostgreSQL · Redis · MongoDB
├── mlops/ MLflow · Weights & Biases · DVC · Optuna
├── infra/ Docker · Kubernetes · GitHub Actions · GCP · AWS
└── learning-now/ quantization · KV-cache · inference serving
