I'm Sushant Satyam — ex-Indian Army officer turned engineer. I build agentic AI, synthetic-data pipelines, and distributed systems designed to perform when failure is not an option.
Ex-Indian Army officer turned engineer. Years leading men, systems, and missions in some of the most demanding environments imaginable — artillery operations, counter-insurgency deployments — where decisions carried real consequences and pressure was a constant companion. Today that same precision fuels my work in Data Engineering, AI, and distributed systems.
I build scalable data platforms, intelligent automation, AI-driven architectures, and high-performance cloud systems — using SQL, Python, PySpark, AWS, Databricks, Kafka, Snowflake, and modern ML frameworks. My work revolves around transforming fragmented data into actionable intelligence, and designing systems that think, learn, and scale.
Beyond engineering, I'm deeply invested in Generative AI, Agentic AI, synthetic data generation, and building technology that augments human decision-making in real-world environments — from SynthForge (LLM-native synthetic tabular data) to CodeWithColonel (a public ML + AI agent curriculum, built in the open).
What the Army taught me still defines me: discipline under chaos. Ownership without excuses. Calmness under pressure. Mission-first thinking. And the ability to lead when the situation is uncertain.
> I don't just build pipelines and platforms. I build systems designed to perform when failure is not an option.
LLM agents, RAG, GraphRAG, tool-use, memory, evals. From SynthForge's LLM-augmented synthesis pipelines to agent architectures that route, plan, and recover — I build AI that performs under load, not just on a demo reel.
Scalable platforms on SQL, Python, PySpark, AWS, Databricks, Kafka, Snowflake. Synthetic generation, privacy intelligence, semantic inference. If the data is wrong, the product is wrong — I spend the weeks upstream so everything downstream tells the truth.
Distributed systems, cloud architecture, full-stack delivery — Python/Next.js/NestJS/Prisma. Built like the Army taught me: discipline under chaos, ownership without excuses, designed to perform when failure is not an option.
A mix of AI/ML systems, trading infra, logistics, and full-stack product work. Open source projects link out to GitHub; private ones are sketched briefly.
Generates high-fidelity synthetic datasets from tiny production samples. Combines Gaussian Copula, CTGAN, TVAE, TabSyn, and Diffusion with LLM-powered semantic inference, PII/MNPI detection, and business-rule extraction. Scales from 1K to 10M+ rows with a built-in eval framework for fidelity, utility, and privacy.
A public, two-track learning repo: six AI/LLM projects covering prompt engineering, agent patterns, tool-routing, memory, and RAG — plus a 21-day ML curriculum from data cleaning to deep learning and recommenders. Progress over perfection.
A working notebook of prompt engineering patterns — structured output, role conditioning, evals, and tool-use — maintained as I iterate on LLM applications.
A TypeScript exercise modeling elastic load-balancer routing semantics. A small, focused repo exploring the edge cases of path-based routing.
Ingests public tender corpora, extracts requirements, and drafts bid outlines with citation-grounded rationale. Built for procurement teams buried in PDFs.
A logistics platform experiment — dispatch, tracking, and ops surfaces for last-mile workflows.
AI-driven optimization for logistics routing and supply-chain flows. Focused on where planners actually spend their time.
Agri-tech product work combining farm data capture with decision-support surfaces for growers and ops teams.
A web product exploring a books-centric UX. Built on the feature/web branch with a TypeScript stack.
A compact Python solution to a take-home engineering challenge — kept private but shippable.
An engine for parsing, analyzing, and surfacing insights from report corpora — the plumbing behind a reporting product.
Third-generation iteration of a long-running personal product line focused on vision / insight tooling.
SQL layer and query patterns supporting the Drishti product experiments.
An older Python sandbox kept for reference — part of how I got here.
// private repos: details intentionally light. happy to walk through any of these over a call.
Small interactive things I've built or sketched — usually as a Claude artifact. Click through to play with them live.
A standalone one-pager demo for TheBookLane, a web product focused on sharing books and growing community.
A personal tracker for 26 foundational AI/ML papers + Google's 5-Day AI Agents Intensive whitepapers — all in one place with checkboxes, topic filters, and direct links to every paper. From Attention Is All You Need → Chinchilla → LLaMA → RoPE → FlashAttention → RAG → InstructGPT → DPO → ReAct → DeepSeek-R1 → Scaling Monosemanticity, plus Google's agent series (Intro to Agents, MCP & Tool Use, Context Engineering, Agent Quality, Prototype to Production). Free, no login.
Soup-to-nuts walkthrough of how agent memory actually works under the hood. Part 1: why agents need memory & the RAG problem statement. Part 2: embedding models (text → vectors, InfoNCE training). Part 3: backpropagation — how embedding weights get learned. Part 4: vector RAG (semantic retrieval). Part 5: Graph RAG & entity extraction. Part 6: the Leiden algorithm for graph communities. Part 7: Microsoft GraphRAG, formalised. Part 8: HNSW — how vector DBs search fast. Part 9: a grand unified architecture connecting every algorithm into one pipeline, with a complete formula reference.
A complete, self-contained dark editorial explainer (Instrument Serif + DM Mono) that builds positional encoding from first principles across 6 sections: metric goals, single-wave ambiguity, unit-circle intuition, sin-vs-sin+cos comparison, full formula breakdown, and an interactive pair explorer with live computed values and wave-speed shifts for i=0→255.
A complete first-principles walkthrough of the full Transformer pipeline with a dark editorial aesthetic: encoder/decoder big picture, embedding geometry, positional encoding, interactive attention visualiser, Add & Norm derivation, FFN intuition, decoder generation stepper, output softmax, and training dynamics with live loss curves.
An interactive browser demo that compares two text inputs using similarity features (Jaccard, fuzzy ratio, n-gram overlap, length, common words) and explains how an XGBoost-style classifier would combine them into a final match score and verdict.
A forensic-style interactive report mapping the full attack chain from compromised GitHub Actions to malicious LiteLLM PyPI releases, suppression botnet behavior, indicators of compromise, and an actionable response checklist for engineering teams.
A visual deep dive into Data Mesh: why centralized data platforms bottleneck at scale, the four core principles (domain ownership, data as a product, self-serve platform, federated governance), and when this paradigm is the right fit.
A deep interactive walkthrough of SynthForge: six synthesis backends, LLM-augmented schema + privacy intelligence, and a five-layer evaluation stack for generating high-fidelity synthetic tabular data from small production samples.
A strategic visual explainer of how software delivery shifts from human handoffs to agent-orchestrated execution: intent synthesis, parallel implementation, sentinel quality gates, MCP-enabled toolchains, and phased enterprise migration.
An interactive research report on modern RAG: adoption trends, eight leading techniques (GraphRAG, Self-RAG, CRAG, Agentic RAG, and more), architecture trade-offs, key papers, enterprise adoption patterns, and persistent risk surfaces.
A live snapshot of the last 365 days on github.com/sushantsatyam. Reflects public commits by default — private repo contributions show up too if enabled in GitHub profile settings.
A working set, not a museum. I'm opinionated about a few things and pragmatic about the rest.
I'm open to interesting problems — AI systems, data platforms, fintech, or anything where the domain is messy and the bar is high.