Who this is for
Working professionals who already know Python, have some mathematical maturity, use AI natively, and feel pressure to transition into serious AI engineering before the market moves past them.
Program update · Live cohorts closed
All live cohorts are officially closed. November 2026 is Ayush Singh’s final live cohort, and enrolment is already closed. No further live cohorts will open.
Ayush is introducing the FDE + Agentic Engineering Program: an annual community membership starting at ₹24,000 + GST. It includes recordings from all three cohorts, a structured learning route, access to Ayush through the community, and one live group Q&A each month.
Every two months, we will add a detailed write-up of a production-grade project. This is a community membership with recordings and a monthly Q&A, rather than a new live teaching cohort.
We will admit up to 50 new members per month. Invitations for the first 50 seats will be emailed in January 2027. Join the waitlist with your name, LinkedIn profile, phone number, and email address to receive a broadcast notification when monthly applications open.
Join the January 2027 waitlist ↗Recorded AI/ML and engineering foundations for working professionals.
For professionals who feel AI is catching up fast, want to upgrade seriously, and need a structured transition into AI-native engineering roles.
This is not prompt engineering, not a passive course, and not a beginner introduction to programming. These recordings cover a dense engineering curriculum for working professionals who want to upgrade into AI engineering, reduce career-risk from AI disruption, build systems, evaluate them, defend decisions, and operate under real constraints.
The public map describes the recorded cohort curriculum. Live cohort enrolment is closed; the new program offers a structured route through the recordings.
Probability, statistics, information theory, linear algebra, optimization, classical ML, deep learning, model evaluation, regularization, gradient flow, and engineering judgment.
Sequence models, attention, GPT internals, RoPE, KV cache, MoE, inference, quantization, LoRA, QLoRA, batching, and serving tradeoffs.
GPU architecture, Flash Attention, FastAPI, async systems, Postgres, Redis, queues, Airflow, Kafka, embedding pipelines, and vector database internals.
Context engineering, structured outputs, memory systems, retrieval science, GraphRAG, tool calling, agent loops, HITL, durable execution, and dynamic workflows.
Golden datasets, slice evaluation, LLM-as-judge, regression gates, tracing, cost dashboards, guardrails, prompt injection defense, red-teaming, and reliability patterns.
Bellman equations, policy gradients, reward modeling, RLHF, DPO, GRPO, RLVR, reasoning models, test-time compute, diffusion, VAEs, and research taste.
A production AI system with architecture decision records, an evaluation harness, deployment, monitoring, cost tracking, a live demo, and public proof-of-work.
Participants first experience the failure, then study the invention, then build, explain, evaluate, and publish.
Participants begin inside a broken system, impossible constraint, or misleading output so they feel the problem before learning the solution.
Older concepts are recalled without notes so understanding compounds instead of disappearing after each week.
The invention is introduced only after the failure is real. Participants watch the thinking, not just the final code.
A scaffold shows how a serious system is structured: interfaces, decisions, tests, logs, and failure cases.
The scaffold disappears. Participants build, debug, evaluate, and defend their own implementation.
If a participant cannot explain the tradeoffs, alternatives, and failure modes, the work is not done.
Learning becomes visible through demos, writeups, technical posts, evaluation reports, and capstone artifacts.
The cohort teaches participants to design the system around the AI, not merely ask an IDE to generate code.
Design bounded build-measure-correct loops where AI generates, tests, reads failures, improves, and stops under explicit success criteria.
Move beyond one-shot prompting into workflows that perceive, plan, act, observe, recover, escalate, and preserve state across long-running tasks.
Use smaller or open-source models for mechanical work while reserving frontier models for reasoning-heavy decisions, synthesis, and judgment.
Design systems where different models play different roles: planner, executor, critic, verifier, summarizer, retrieval judge, or domain specialist.
Coordinate subagents, tools, queues, retries, memory, logs, and human approvals so the system behaves coherently under changing context.
Add permission boundaries, policy checks, destructive-action approvals, human review points, fallback paths, and audit trails for serious production work.
Serious agents do not operate in clean tutorial environments. They face partial information, missing capabilities, ambiguous requests, strict policies, token budgets, latency limits, and evaluators that punish inconsistent behavior.
The point is not whether a model can answer once. The point is whether a system behaves reliably when the environment is constrained, noisy, and adversarial.
The sprint is built to create engineering depth, research taste, proof-of-work, and opportunity creation at the same time.
Coding agents, AI IDE workflows, terminal-first development, code review, repo understanding, patch verification, and AI-generated code safety.
Read, critique, reproduce the core idea, present the insight, and defend what aged well or failed.
Study real AI failures: hallucination, tokenization, prompt injection, silent model failure, RAG failure, reward hacking, and cost blowups.
Every week participants defend a real decision: what they would do, why, what tradeoff they make, and what alternative they reject.
Ship, talk to users, publish proof-of-work, understand markets, build outreach systems, and create opportunity instead of waiting for it.
This page shows the curriculum architecture. The full syllabus includes each week's topics, papers, productive failure trigger, build work, self-explanation checkpoint, interview mapping, and deliverables.