Forward Deployed Engineer at Im Beside You (Tokyo). I architect on-chain data pipelines (Airflow / dbt / BigQuery + web3.py) and ship LLM agents with tool use, structured outputs, and ReAct loops. Same engineer designs the warehouse, writes the decoder, builds the agent, and wires the UI.
I'm a software engineer who turned into an agentic AI engineer the moment the tools became good enough to be worth shipping on. I build production systems where LLM agents decide, retrieve, and act — not chat surfaces.
Tokyo-based AI startup. Sole engineer on agentic infrastructure. I'm shipping ProcMine — a fully autonomous process-mining platform where LLM agents handle unstructured activity classification while RL agents in Gymnasium environments learn the segmentation policy. Replacing months of consulting with a runtime.
No multi-step command runs without the agent restating it first. No “deploy succeeded” without user-visible artefact verification. Every correction in-session becomes a [LEARN] rule indexed in SQLite — the same mistake never costs two corrections.
Claude Code (Opus 4.7, 1M ctx) as primary builder, Codex CLI as adversarial reviewer + data-prep specialist. Iterative critique loop up to 5 rounds. Custom split-memory (CLAUDE / AGENTS / SOUL / LEARNED), 8 hooks, 9 slash commands, 3 specialized subagents. I dogfood it daily.
Graduated May 2022. Engineering physics taught the habit I lean on most: derive from first principles, then optimize. It maps cleanly onto agent design.
Agentic enterprise process mining · Im Beside You
Fully autonomous platform where LLM agents discover, segment, and optimize enterprise workflows. RL agents operate in Gymnasium environments modelled on real process traces, learning segmentation policies through reward signals. 13-step multimodal pipeline behind nginx, in production.
On-chain ABI-driven parser · KLab engagement
ABI-driven decoder (Python + web3.py) parsing raw EVM logs into typed records. Handles ERC-20, ERC-721, Uniswap V2/V3, custom contract events. Reorg-aware backfills, incremental loads into BigQuery for tokenomics analytics. BTC on-chain metrics (MVRV, NVT, hash rate, UTXO) aligned alongside derivatives + macro.
Trading data infrastructure · FlexTrade Systems
Large-scale ingestion + analytics platform processing institutional order flow across asset classes (equities, FX, fixed income, crypto). Airflow + dbt pipelines over multi-billion-row warehouses, sub-100ms query layer. 3× throughput, −50% cloud spend. Crypto/on-chain extension via web3.py.
Dataset-generation SDK · YAML DSL → AST compiler
A handful of YAML config lines fans out into 20,000+ synthesized datasets / strategies. Custom DSL → restricted-Python AST compiler evaluated over multi-source data (on-chain, derivatives, macro). Monte Carlo + Combinatorial Purged CV filter survivors. Same engine drives parameter sweeps and eval harnesses.
Autonomous trading-signal pipeline
ReAct-style agent that ingests financial news + social signal, plans its own retrieval, classifies sentiment via structured JSON schemas, and emits actionable alpha signals. Tool-use with confidence-based routing, retry, and fallback strategies wired in.
Mamba + LSTM + HMM market-state classifier
Time-series ensemble over BTC OHLCV and on-chain metrics, exposed as a callable agent tool. Orchestrator agents invoke it for real-time market state classification; paper-trading harness evaluates regime-conditional strategies autonomously.
Parallel-batched reads, schema gating, subagent context isolation. Compounding latency wins across 40-turn sessions.
ABI-driven decoder over raw chain logs. Reorg-aware backfills, incremental loads into BigQuery, typed records for tokenomics.
RL agent learning segmentation policy on synthetic process traces. Reward shaped by downstream conformance, not heuristics.
If you’re building production agentic systems, let’s talk.
Neeshant Pandey is a full-stack AI engineer based in Pune, India, currently working remotely as a Forward Deployed Engineer at Im Beside You (Tokyo). He graduated from IIT Hyderabad with a B.Tech in Engineering Physics in 2022 and has over four years of production engineering experience across agentic AI systems, LLM pipelines, data infrastructure, and full-stack development.
Neeshant builds the full AI stack end-to-end: LLM agent systems (ReAct loops, tool use, structured outputs, agentic RAG), on-chain data pipelines (Airflow, dbt, BigQuery, web3.py), ML systems (PyTorch, Mamba, LSTM, reinforcement learning via Gymnasium), and full-stack web (FastAPI, Node.js, React, Next.js, AWS, GCP, Docker). He is the sole engineer on agentic infrastructure at Im Beside You, shipping ProcMine — a fully autonomous enterprise process mining platform.
ProcMine is an agentic enterprise process mining platform built at Im Beside You. LLM agents discover, segment, and optimize enterprise workflows; reinforcement learning agents operate in Gymnasium environments modelled on real process traces, learning segmentation policies through reward signals. The 13-step multimodal pipeline runs in production behind nginx.
Neeshant runs Claude Code and OpenAI Codex CLI in parallel as a two-agent development system. Claude is the primary builder (architecture, implementation, file edits); Codex acts as the adversarial reviewer and data-prep specialist. The two agents are wired into an iterative critique loop up to five rounds until both agree. He has built a custom split-memory architecture (CLAUDE.md, AGENTS.md, SOUL.md, LEARNED.md), an FTS5 SQLite learning log, custom slash commands, hook-enforced quality gates, and three specialized subagents (scout, planner, reviewer).
Yes. Neeshant is open to senior or staff-level roles in AI engineering, data engineering, or generalist software engineering — remote or based in San Francisco, Bangalore, or Tokyo. The fastest way to reach him is by email at [email protected].
Neeshant earned a B.Tech in Engineering Physics from the Indian Institute of Technology (IIT) Hyderabad, graduating in May 2022.
Python (primary), TypeScript, JavaScript, SQL, Bash. AI & LLM: Anthropic Claude API, OpenAI API, ReAct, tool use, structured outputs, MCP, agentic RAG, agent evaluation. ML: PyTorch, Mamba, LSTM, HMM, reinforcement learning (Gymnasium), Monte Carlo, Combinatorial Purged CV. Backend: FastAPI, Node.js, REST APIs, PostgreSQL, Redis. Frontend: React, Next.js, Tailwind. Data: Apache Airflow, dbt, BigQuery, Pub/Sub, web3.py, ABI-driven EVM event decoding. Cloud: AWS (EC2, S3, Lambda), GCP, Docker, Terraform, CI/CD.
Email: [email protected]. GitHub: github.com/neeshant-pandey. LinkedIn: linkedin.com/in/neeshantpandey. X: @1neeshantpandey.