One month of Applied AI training, taught by professionals from hiring companies, followed by a placement drive. Your fee is based on your assessment score.
Start Your Application →Fill out the form below with your details.
Complete a 20-minute preliminary assessment to unlock your fee waiver percentage.
Pay your assessed fee and enroll in the upcoming cohort.
60 hours of Applied AI training over one month, led by industry professionals.
DiscoveHr Pro unlocks — apply directly to eligible jobs.
Takes about 5 minutes. You'll receive your assessment link by email right after.
A project-based curriculum where every team builds one end-to-end AI system across a month of live weekend classes and weekly team build hours — guided by faculty acting as the team's technical lead.
Each week opens with live teaching, then hands the project to the team for the build. Teams select a capstone track in Week 1 and carry it through to a live Demo Day at the end of Week 4.
requirements.txt are ready before Week 1 opens, so no team loses build time to environment setup.| Day | Block | Format | Session content |
|---|---|---|---|
| Sat | 1 (90m) | Keynote + framing | The expert mental model of modern AI; production vs. academic AI. Project briefing, team formation, and capstone track selection. |
| Sat | 2 (90m) | Deep dive | Transformer internals beyond the tutorial: Flash Attention 2 memory budgeting, RoPE/ALiBi, KV cache mechanics, GQA/MQA, MoE routing — including a by-hand RoPE rotation exercise and a KV cache-size derivation. |
| Sun | 1 (90m) | Deep dive | Multimodal architectures: CLIP/InfoNCE contrastive pre-training, ViT patching vs. CNN features, cross-modal fusion strategies, diffusion U-Net internals. |
| Sun | 2 (90m) | Live-coded walkthrough | The TL live-codes the project scaffold end-to-end — repo + CI, attention profiling benchmark (naive vs. Flash Attention) — so the team starts activity hours with a working baseline. |
| Day | Hours | Team activity |
|---|---|---|
| Mon | 2h | Team clones the scaffold, confirms the CI pipeline runs green, assigns the five project roles. |
| Tue | 2h | Run the attention profiling benchmark at sequence lengths 512–4096; capture the naive vs. Flash Attention chart. |
| Wed | 1h + 1h | 1h TL office hour for blockers · 1h begin data pipeline design for the chosen project track. |
| Thu | 2h | Continue the data pipeline build; draft the team's architecture diagram. |
| Fri | 1h | Finalize the architecture diagram; post the Week 1 deliverable to the shared repo. |
Repo + CI initialized · data pipeline designed · attention profiling chart (naive vs. Flash Attention) · team architecture diagram.
| Day | Block | Format | Session content |
|---|---|---|---|
| Sat | 1 (90m) | Deep dive | Parameter-efficient fine-tuning: LoRA/QLoRA/DoRA math and trade-offs, rank selection strategy, adapter merging (SLERP/TIES/DARE), quantization (GPTQ/AWQ/NF4). |
| Sat | 2 (90m) | Live-coded walkthrough | The TL live-codes a QLoRA fine-tune on a shared Unsloth notebook (fits a single T4 GPU) so every team has a working baseline before touching their own data. |
| Sun | 1 (90m) | Deep dive | Alignment: the RLHF three-model system and PPO instability, the DPO derivation that removes the reward model, SimPO/KTO/IPO, Constitutional AI. |
| Sun | 2 (90m) | Deep dive + walkthrough | Evaluation beyond perplexity/BLEU: LLM-as-judge (G-Eval/MT-Bench), contamination and Goodhart's law, red-teaming — plus a live walkthrough of building a DPO preference dataset from an instruction set. |
| Day | Hours | Team activity |
|---|---|---|
| Mon | 2h | QLoRA fine-tune the project's base model on domain data; log rank/loss experiments. |
| Tue | 2h | Build 100+ DPO preference pairs (chosen/rejected) specific to the project track's failure modes. |
| Wed | 1h + 1h | 1h TL office hour · 1h run the DPO training loop (TRL) on the preference set. |
| Thu | 2h | Build the LLM-as-judge evaluation pipeline; run the custom eval suite against the tuned adapter. |
| Fri | 1h | Integrate the tuned checkpoint into the project backend via a FastAPI /generate endpoint. |
Fine-tuned adapter checkpoint · DPO preference dataset (100+ pairs) · custom evaluation report · integrated /generate endpoint.
| Day | Block | Format | Session content |
|---|---|---|---|
| Sat | 1 (90m) | Deep dive | Advanced RAG: naive RAG failure modes, semantic/hierarchical chunking, hybrid dense+BM25 search with Reciprocal Rank Fusion, HyDE, RAGAS metrics (faithfulness, answer relevance, context recall). |
| Sat | 2 (90m) | Live-coded walkthrough | The TL builds a hybrid-search RAG pipeline with a RAGAS evaluation harness end-to-end on a shared dataset. |
| Sun | 1 (90m) | Deep dive | Agentic AI: ReAct / Chain-of-Thought / Tree-of-Thought reasoning, function calling and structured output, memory tiers (in-context, vector store, episodic), LangGraph / CrewAI / AutoGen. |
| Sun | 2 (90m) | Deep dive + live build | Compound AI systems: the compound-AI thesis, orchestrator-worker patterns, DSPy prompt optimization, guardrails (Outlines/NeMo) — plus a live build of a plan → execute → reflect LangGraph agent. |
| Day | Hours | Team activity |
|---|---|---|
| Mon | 2h | Implement hybrid search over the project's corpus; wire the chunking strategy. |
| Tue | 2h | Run RAGAS on 10+ hand-crafted QA pairs; fix retrieval before touching generation if faithfulness falls below 0.70. |
| Wed | 1h + 1h | 1h TL office hour · 1h build the LangGraph agent with its first tool. |
| Thu | 2h | Add remaining tools (target 3+); wire the agent into the project interface. |
| Fri | 1h | Run the first end-to-end query through the full pipeline; log the RAGAS evaluation report. |
Working RAG pipeline with RAGAS scores · functional agent with 3+ tools · first end-to-end demo · RAGAS evaluation report.
| Day | Block | Format | Session content |
|---|---|---|---|
| Sat | 1 (90m) | Deep dive | Inference engineering: speculative decoding, vLLM / PagedAttention, INT4/INT8 quantization trade-offs, distillation, token cost modelling (the API-vs-self-host break-even exercise). |
| Sat | 2 (90m) | Deep dive | LLMOps and safety: experiment tracking and monitoring (W&B / Langfuse / DVC), failure taxonomy (hallucination, prompt injection, jailbreaks, bias), bias auditing, EU AI Act overview. |
| Sun | 1 (90m) | Demo Day | Team presentations — 10 minutes each (problem, architecture, live demo, metrics/learnings) plus 5 minutes Q&A. Judging: technical depth (30%), working demo (30%), evaluation rigour (20%), presentation clarity (20%). Scale block length to team count. |
| Sun | 2 (90m) | Frontier + closing | Lightning session on test-time compute / reasoning models and world models / Mamba, followed by awards and a closing session on research and engineering career pathways. |
| Day | Hours | Team activity |
|---|---|---|
| Mon | 2h | Deploy the project via vLLM behind FastAPI; confirm the OpenAI-compatible endpoint responds. |
| Tue | 2h | Instrument Langfuse for prompt logging and latency tracking; verify events appear in the dashboard. |
| Wed | 1h + 1h | 1h TL office hour · 1h build the hallucination monitor (LLM-as-judge on sampled outputs) and add the guardrail layer. |
| Thu | 2h | Complete the safety audit checklist (bias audit, red-team log); draft the 4-page technical writeup. |
| Fri | 1h | Rehearse the 10-minute Demo Day presentation live on camera at least once. |
Live deployed endpoint · monitoring dashboard (latency, cost, hallucination rate) · safety audit checklist · 4-page technical writeup · GitHub repo with README and evaluation results · Demo Day presentation.