Hitansh Gopani · he/him · Mumbai, India
AI/ML Engineer @ Pitch Perfekt Collective · K J Somaiya Institute of Technology (IT Engineering)
linktr.ee/Crypto_HG · linkedin.com/in/hitanshgopani · github.com/Hitanshuser50
I build AI systems people can rely on every day, not just clever demos. Most of my work sits where a model's output has to be checked before anyone acts on it: voice intake that turns answers into validated structured data, ML pipelines with honest evaluation, and on-chain systems where the claim is verifiable instead of taken on trust.
Currently an AI/ML Engineer at Pitch Perfekt Collective, working on the systems that keep creative AI workflows organized, reusable, and production-ready.
Voice interviews for regulated intake, where every answer has to be defensible.
- Questions are read verbatim, and spoken answers become checked, structured data rather than free-form transcripts.
- Produces an explainable risk score, then routes the case to a human reviewer with a full audit trail.
- Ships nine use-case packs, each with its own per-agent voice.
Battleship on Midnight where hidden fleets are enforced by zero-knowledge proofs, not by a trusted server.
- Each fleet is committed on-chain as
hash(board, salt); every hit/miss answer is a proof checked against that commitment, so moving ships mid-game fails verification. - Compact contract with six circuits, deployed to Midnight Preprod; squad mode runs a two-player game entirely on-chain through the Lace wallet.
- 267 automated tests, including a 54-case adversarial suite that plays the cheater; deterministic three-tier AI opponent that plays through the real game API.
Verifiable fine-tuning on 0G: every fine-tuned model gets a passport anyone can check without a wallet.
- Collects a model's lineage (base-model hash, dataset storage root, hyperparameters, TEE-attested delivery) into a manifest on 0G Storage, with its
keccak256anchored by a verifiedPassport.solcontract. - Diagnosed an SDK retrieval defect that lost a paid fine-tune on Windows, then replaced the download path with an HTTP retriever that re-derives the storage root before acknowledging.
- 104 contract tests; three passports minted on 0G Galileo testnet, the latest from a real adapter with a verified attestation.
A fair-value layer for dreamDEX event contracts on Somnia: what the odds should be, next to what the market quotes.
- Computes fair probability with
Φ(d₂)in Solidity from on-chain EWMA realized volatility, publishing Gaussian and Student-t estimates side by side with edge, break-even, and Kelly sizing. - Five contracts deployed and verified on Somnia Shannon testnet, with 117 Hardhat tests and math cross-checked against SciPy.
- Includes an Edge Radar frontend, a backtesting harness, and a trading bot (dry-run validated).
One workspace for medical insurance claims across patients, hospitals, insurers, and admins.
- Role-based dashboards covering the full claim lifecycle from submission to settlement, with document handling and analytics.
- Gemini-backed assistant via the Vercel AI SDK, plus a coverage calculator for out-of-pocket estimates.
- Smart-contract settlement examples in Solidity and Rust; built on Next.js App Router, TypeScript, Tailwind, and shadcn/ui.
Link prediction and unsupervised fraud detection over the Ethereum transaction graph.
- GraphSAGE learns a 64-dimensional embedding per wallet; pair scores predict future transactions and an Isolation Forest flags anomalous wallets without labels.
- Training and serving are split at an artifact boundary, so the Streamlit dashboard never imports PyTorch and runs on a free CPU tier.
- Live mainnet lookup through Etherscan and Web3.py scores real addresses against the trained model.
| Project | What it is |
|---|---|
| incident-intel | Video anomaly detection as a running system: versioned data pipeline, three-way model comparison on UCSD Ped2, PyTorch autoencoder, FastAPI backend, and a Next.js + Three.js dashboard. |
| trustllm | LLM evaluation and governance platform: run a prompt across models, compare metrics, and gate deployment with configurable guardrails and audit logs. |
| SapnaForge | Multilingual pipeline that turns handwritten and spoken business ideas into structured pitches with feedback; built for Code Odyssey 4.0 with Tata STRIVE. |
ML: PyTorch, GraphSAGE, scikit-learn, OpenCV · LLMs: Gemini, LangChain, Vercel AI SDK · Chains: Midnight (Compact), 0G, Somnia, Ethereum · Data: Snowflake Cortex, SQLite, MongoDB
Also an AI/ML Trainer at QUASTECH and a Google Arcade Facilitator, teaching cloud and GenAI to students who are usually earlier in the journey than I am.
- AI/ML Engineer, Pitch Perfekt Collective — building the systems that keep creative AI workflows organized, reusable, and production-ready. post
- Snowflake "Zero to Agents" — Cortex Analyst (text-to-SQL), Cortex Search (RAG), semantic layer, dynamic tables, and column-level governance, end to end in one session. post
- AI/ML Trainer, QUASTECH · Zonal Winner, Avishkar · GitHub Universe Mumbai recap — all from one December. post
- Global Fintech Fest 2025 — tokenization, agentic payments, and India's real-time payments scale, from the floor. post
- 1st Runner-Up, Code Odyssey 4.0 (Tata STRIVE × K J Somaiya Institute of Technology) — OCR + multilingual NLP + LLMs to turn handwritten business ideas into structured pitches. post
- Snowflake World Tour, Mumbai — data foundations, governed AI at scale. post
- Google Arcade 2025 Facilitator — ran a mentorship group that hit a 94% completion rate against a 22% overall average. post
- AWS Summit Mumbai 2025 · Lean Six Sigma Foundations + Apply Lean Six Sigma to Services certifications. post
Snowflake Zero-to-Agents · Lean Six Sigma (Foundations + Services) · Zonal Winner, Avishkar · 1st Runner-Up, Code Odyssey 4.0
linktr.ee/Crypto_HG · linkedin.com/in/hitanshgopani · github.com/Hitanshuser50