Curriculum vitae
Kanishk Paul. Kolkata.
Research interests
Rigorous evaluation of AI systems; world models and causal induction in interactive agents; multilingual and low-resource NLP; efficient inference under tight compute.
Education
B.Tech, Computer Science and Engineering (Internet of Things and Cyber Security including Block Chain Technology), Heritage Institute of Technology, Kolkata.
National Gems Higher Secondary School, Kolkata.
Research experience
Jadavpur University, Kolkata. Research collaborator, multilingual NLP. Co-authoring a manuscript (in preparation) with Prof. Sudip Naskar and a TU Munich collaborator.
- Built a reproducible round-trip evaluation (14 Indic languages, two translation-model families, 56,000 translations) measuring whether honorific register survives a pivot through English, on 4-bit local models and free-tier GPUs.
- Identified a systematic deference floor: IndicTrans2 kept 99% of deferential source register (1,659/1,677) but 6% of less-deferential register (71/1,188) among classified outputs; native-speaker validation of cue tables for 12 of 14 languages is pending.
- Showed that dedicated MT's near-zero pivot penalty is a floor artefact, since its direct path already drops almost all less-deferential register, while LLM translators lose 20–22 points through their own English pivot.
- Wrote LLM-free, rule-based register extractors (one cue table per language) and a Unicode normalization policy that fixed four silent under-counting bugs; froze datasets behind hash checks and made every pipeline resumable and seeded.
- Preregistered predictions before each phase and reported two that failed: the original contextual-fraction predictor did not survive robustness checks, and the target-decoder-prior test failed with a flawed metric; a corrected confirmatory re-test is preregistered.
Manuscripts in preparation
Register obstruction, with S. Naskar and a TU Munich collaborator. In preparation.
Mechanic Composition Worlds. In preparation.
Research projects
World models for ARC-AGI-3: Virgil and its predecessors, with every score and its rules.
ButterflyGate: an O(n log n) token mixer in a pretrained model's early layers.
ResonatorLM audit: an independent replication.
Earlier: NeuroSymbolic Chain of Thought, Synthesis of Thought, Debate of Thought, a photonic Shor's algorithm design, an MCMC sampler, a fee-less AI-miner blockchain, federated learning with proof-of-work. See History.
Industry and other
Unify (YC W23). Design partner. Tested a pre-release agent platform and reported reproducible gaps in OAuth permission visibility, learned-memory review, action audit trails and per-action cost attribution.
Heritage Institute of Technology. Undergraduate researcher, humanoid robotics. Wrote and defended a proposal for a low-budget assistive humanoid robot through several faculty review rounds; informal funding commitment of ₹20,000 from the department.
buildspace, Season 5. Built an offline LLM for Android that acts (messages, calls, music) solo in about three weeks.
Selected projects
Blindfold. A browser agent that redacts pages on device and sends only stand-in values through one audited exit. Co-led a six-person SIH 2026 team.
maiku. One runtime for Python functions, local LLMs, PyTorch modules, MCP servers, A2A agents, CLIs and people, through one streaming event contract.
ChromeClaw. A browser agent with one schema-validated action per step, a policy layer and a JSONL trace of every step.
ReasonTrace. A visual debugger that turns agent logs into a seven-event graph.
Sefai. Offline GGUF inference in about 150 lines of Rust.
Also: Casca, a desktop agent runner. Stockfih, Stockfish 18 in the browser. ParaReport, Kolkata civic reports. CuteTuner, a local vocal tuning studio. lifesim, a Monte Carlo life simulator. Horizon, a local decision workstation.
Earlier builds, 2020 to 2025: an answer engine, English to SQL, no-code fine-tuning, an AI music critic, a map you talk to, a speaking coach, flood and drought mapping, a deepfake detector, a university advisor, a disaster-response map, a lunar game, a multimodal assistant, trading systems. See Work.
Ventures
OpenMind. CEO.
faculza.
NashBase, with a potential co-founder.
A research collective. Wound down.
A college AI research group.
Hackathons
8 hackathon wins.
| Year | Event | Field | Result | Built |
|---|---|---|---|---|
| 2023 | SIH 2023, college round | Runner-up, cash prize | Disaster-response platform; led the AI layer | |
| 2023 | NASA Space Apps | Unity lunar exploration module | ||
| 2024 | SIH 2024, college round | Top of the pitch round | Deepfake image detector | |
| 2024 | NSHM Innovathon | 100+ teams | 3rd, behind IIT Kharagpur and IIT Kanpur | University advisor; owned the AI |
| 2024 | MLX (GDG On Campus, inter-college, Kaggle) | ~90 | 8th; led both tracks for long stretches | Classification and regression |
| 2025 | MLX 2.0 (GDG On Campus, Kaggle) | ~50 | 1st | Classification and regression |
| 2025 | National hackathon with investor scouting (up to ₹1 crore) | Cleared the pitch round; 23rd, top 21 advanced | AI music critic, built solo | |
| 2025 | Smart Bengal Hackathon | ~1,100 teams | Finalist (top ~80); about top 20 | Speaking coach; led the AI |
| 2025 | ISRO Bharatiya Antariksh Hackathon | Did not reach the final | Flood and drought spread | |
| 2025 | SIH 2025, college qualifier | Won the pitch round; qualified for nationals | City traffic platform; led the team | |
| 2025 | Hack Heritage 2025 | ~300 teams | Won the pitch round; selected for SIH nationals | Ecology map platform, built solo in ~12 h |
| 2026 | SIH 2026 | Blindfold; co-led, most commits | ||
| 2026 | Build with Gemma Hackathon (virtual, 4 Jul) | ParaReport | ||
| 2026 | Amazon ML Challenge 2026, Business Entity Resolution | top team 0.991483; 10th 0.990282 | 0.988216 macro F0.5 | Entity matching across 2.2M businesses |
| 2026 | ARC Prize 2026 (Kaggle) | Account best 0.23 | Blackstar |
Honors and selections
Google × Hugging Face Gemma meetup, 1 of 50.
DisCo Cohort 001.
PrettiFlow Cohort 2, selected 1 of 6, declined.
Thinking About Thinking Ambassador Programme.
F Inc Blueprint 2, interviewed.
Entrepreneurs First Fall 2026, interviewed.
Invited as a design partner by a startup from DeepMind founders.
Open source
unifyai/unify-agent#171, merged.
nossa-y/activity-frames#47, open.
Skills
Languages: Python, TypeScript, JavaScript, Rust, C++, SQL, Java.
ML and AI: PyTorch, TensorFlow, Hugging Face Transformers and Hub, scikit-learn, LLMs (inference, fine-tuning, evaluation), RAG, prompt engineering, machine translation (IndicTrans2, NLLB-200, FLORES-200), 4-bit quantization, state-space and post-transformer models (Mamba, RWKV, HRMs), neurosymbolic and symbolic AI, reinforcement learning, computer vision, multi-agent systems, multilingual and low-resource NLP, MLX, llama.cpp, ONNX Runtime.
Data: NumPy, pandas, SciPy, NetworkX, linear programming, benchmarking, data pipelines.
Engineering: React, Next.js, Tailwind CSS, shadcn/ui, FastAPI, Flask, Django, Node.js, Supabase and PostgreSQL, Tauri, Playwright, Streamlit, Docker, REST APIs, CI/CD, MLOps, AWS and GCP, Git, Linux, pytest, Vitest.
Practice: experimental design, preregistration, benchmark construction, paper reproduction.