A self-paced, online specialist programme from Bodhih for engineers who already build with LLMs. You extend one reference system through five modules, then build your own over six weeks and defend it to two evaluators. About 45 hours over 10 weeks.
The Bodhih LLM Engineering certification, formally Certified AI/ML Specialist — LLM Engineering (Level 3), is a self-paced online course from Bodhih Training Solutions for engineers who already build with Python. It takes about 45 hours over 10 weeks across 15 stages: a diagnostic, five modules on retrieval, agents, evaluation, production and governance, a six-week capstone with five checkpoints, a recorded viva with two evaluators, a mock and a certification exam. It costs ₹64,999 plus GST.
Your retrieval pipeline works on the questions you tried. When someone asks why that chunking, that embedding model or that value of k, the truthful answer is that it was the default in the tutorial.
Every headline metric holds steady, yet a user sends you an answer with a valid citation and the wrong conclusion. You have an evaluation script. You do not have a gate that fails when it should.
What happens when a retrieved document carries an instruction? What does this cost per thousand questions at real volume? Where do the embeddings live? Each question is fair, and each currently gets a shrug.
You read and write Python confidently and have put a RAG assistant or an agent in front of users. You want the engineering discipline that the first version skipped.
You hold the Certified Applied AI/ML Practitioner (Level 2) credential or have equivalent experience, and you want depth in retrieval, agents and evaluation rather than another overview.
You review LLM system designs and need a firm basis for the questions about latency budgets, cost ceilings, caching risk, rollback and audit trails.
You need evidence, not attendance. Each engineer you sponsor produces a working system, an evaluation report and a governance pack, marked by two people and defended in a recorded viva.
If you have not yet written Python that someone else depends on, this is the wrong starting point: begin with Certified AI/ML Foundations (Level 1) or Certified Applied AI/ML Practitioner (Level 2).
A course records that you watched. This records that your work changed. Imagine handing your engineering manager an evaluation report and a rollout plan instead of a demo. These are the artefacts the stage list produces, each one from your own capstone system.
At least three chunking configurations compared on the same question set by recall@k and context cost, with the choice and its reason written down, including the case where they are within noise and you chose the cheapest.
An agent loop with tool schemas, defined tool-failure and termination behaviour, a guardrail layer, and a table of at least eight prompt-injection attacks against your own system with what changed as a result.
Sixty or more questions with their growth story, retrieval and answer metrics reported separately, evidence that any LLM-as-judge agrees with your own labels, and an error analysis that names failure classes.
A system that starts with one command, a latency measurement under a stated load, a cost model in rupees, a caching decision either way, structured traces, prompt and model versioning, and a rollback story.
A model card for the system, a data-flow and residency note, PII handling, an audit trail, retention, an incident runbook with one worked incident, and a human-oversight statement.
Your system with an 8 to 12 page report, double-marked on five criteria. A third marker is brought in when the two markers disagree on the result.
Thirty minutes with two evaluators: you walk through the system, answer questions on it, and say what you would change and why. Both evaluators must agree that it meets the standard.
Issued when the exam, the capstone and the viva are all passed. It is verifiable, is also issued as an Open Badge 2.0, and is followed by a re-sit of the diagnostic at 90 days.
The stages run in order. Five modules extend one reference system, so each idea is tested on code you already know, and each module feeds a capstone checkpoint. As you read the stage list, notice which of these decisions your current system made by measurement and which by habit.
Measured before you start and again on the same instrument at the end, so the change is recorded rather than claimed.
A 40-minute diagnostic: 30 scenario-led questions across system design and agents, evaluation, production and governance. Not scored for the credential; it sets your baseline and is re-sat at 90 days.
Retrieval architectures, chunking strategies compared by measurement on the 48-document Level 3 corpus, embeddings versus lexical retrieval, hybrid search, re-ranking, context assembly under a token budget, memory patterns and their failure modes. Eight runnable Python cells in the browser extending the reference system. Build: the reference retriever with a measured chunking decision, feeding capstone checkpoint 2.
Planning loops taught before the jargon, function calling and tool schemas, argument validation, state, guardrails (allowed-tools list, step budget, PII scrub, out-of-scope refusal), sandboxing, the failure taxonomy, and a live demonstration of prompt injection arriving through retrieved text. Build: the three-tool agent with a guardrail layer and an injection test, feeding capstone checkpoint 3.
Golden sets and how to grow one from production traffic without leaking, offline eval suites, metrics and what each one hides, human-in-the-loop labelling and inter-annotator agreement, LLM-as-judge validity and its biases, red-teaming, and regression gates in CI — including the silent regression where every headline metric holds steady while the system starts giving wrong answers with valid citations. Build: the eval suite and a CI gate that actually fails, feeding capstone checkpoint 4.
Latency budgets decomposed across the request path, cost modelling in rupees at realistic volumes, caching (exact, normalised, semantic) and the correctness risk of each, streaming and perceived latency, observability and tracing, safety filters, prompt and model versioning, canary rollout and rollback. Build: instrument the reference system, load-test it in the browser and write the rollout plan, feeding capstone checkpoint 5.
Data residency across documents, embeddings and the prompt; PII pipelines for Indian data and the false-positive cost of redaction; audit trails an auditor can use; model cards for systems rather than models; incident response; the regulatory landscape in plain English and explicitly not as legal advice; vendor procurement questions. Build: the governance pack for the reference system, feeding capstone checkpoint 5.
Due Day 32. The problem in one paragraph, the corpus with its licence or sponsor confirmation and anonymisation approach, success criteria as numbers with reasons (recall@k, answer accuracy, p50 and p95 latency, cost ceiling per 1,000 questions, acceptable refusal rate), a risk list and a six-week plan naming what will not be attempted. SME steers in the comment thread within 3 working days. Required; not marked.
Due Day 39. A working retriever and a measurement rather than an opinion: at least three chunking configurations evaluated on the same question set with recall@k and context cost, the choice and its reason (including 'they are within noise, so I chose the cheapest'), embedding-model choice with a reason, the handling of near-duplicates, and a first golden set of at least 20 questions with expected source documents. Required; not marked.
Due Day 46. The agent loop with its tool-failure and termination behaviour, two or more tools with schemas, the guardrail layer, and a prompt-injection test actually run: at least eight attacks against the learner's own system, a table of what happened, and what changed as a result. Attacks that succeeded earn more than attacks never tried. Required; not marked.
Due Day 53. A golden set of 60 or more questions with its growth story, retrieval and answer metrics reported separately, evidence that any LLM-as-judge agrees with the learner on at least 30 hand-labelled items, and an error analysis naming failure classes rather than examples. Required; not marked.
Due Day 60. A running system with one-command startup, a latency measurement under a stated load, a cost model in rupees, a caching decision either way, structured traces, prompt and model versioning, a rollback story — and the governance pack: model card, data-flow and residency note, PII handling, audit trail, retention, an incident runbook with one worked incident, and a human-oversight statement. Required; not marked.
Due Day 67. A deployed or deployable LLM system over the learner's own corpus or the supplied 48-document Level 3 corpus, with an 8–12 page report, the evaluation report, the governance pack, a model card and the AI-use declaration. Five criteria x 0–4; pass at 15/20 or above with no criterion below 2; double-marked, with a third marker when the totals differ by more than 3 or the pass/fail decision differs. One resubmission within 10 working days. Graded within 10 working days. Book the viva slot in this stage.
Thirty minutes, recorded, two evaluators: 5 minutes walk-through, 15 minutes questions from the 40-prompt bank tagged to the five criteria, 10 minutes 'what would you change and why'. Scored meets / does not meet per evaluator with written reasons; both must agree, and disagreement goes to a third evaluator. Held within 2 weeks of the capstone pass. The viva also confirms authorship.
Forty scenario-led questions, sixty minutes, including architecture- and code-reading. Drawn from the Level 3 practice pool; returns per-domain readings and the lessons to revisit.
Sixty scenario-led questions, ninety minutes, under exam conditions, drawn from the sealed Level 3 pool. Pass: 70% overall and at least 60% in every domain. The credential also needs the capstone at 75% or more and a viva judged to meet the standard by both evaluators. One free retake after seven days.
Finishing this pathway issues a credential backed by the evidence you produced along the way. It is verifiable by link and is also issued as an Open Badge 2.0, so an employer can check it without taking your word for it.
Chunking strategies compared on a fixed corpus, embeddings versus lexical retrieval, hybrid search, re-ranking, and context assembly under a token budget. The work a RAG course usually leaves as an exercise.
Planning loops taught before the jargon, function calling and tool schemas, argument validation, state, sandboxing, and a taxonomy of how agents fail.
An allowed-tools list, a step budget, PII scrubbing and out-of-scope refusal, tested against a demonstration of injection arriving through retrieved text.
Golden sets grown from production traffic without leaking, what each metric hides, inter-annotator agreement, the validity and biases of LLM-as-judge, red-teaming, and regression gates in CI.
Latency budgets decomposed across the request path, cost modelling in rupees at realistic volumes, exact, normalised and semantic caching with the correctness risk of each, tracing, canary rollout and rollback.
Data residency across documents, embeddings and the prompt, PII pipelines for Indian data, audit trails, model cards for systems, incident response and vendor procurement questions. Regulation is covered at principle level and is not legal advice.
Want the detail before you decide? Read the guide: LLM engineering: what RAG, agents and evaluation actually require in production.
| What is compared | A typical AI course | LLM Engineering at Bodhih |
|---|---|---|
| What you build | A tutorial assistant over sample documents, following the instructor’s steps. | One reference system extended across five modules, then your own system over your own corpus or a supplied 48-document corpus. |
| How design choices are made | Chunk size, embedding model and prompt are given to you. | You compare configurations on the same question set and write down the choice and its reason. |
| Evaluation | A handful of spot-check questions at the end. | A golden set of 60 or more questions, retrieval and answer metrics reported separately, a validated judge and a CI gate that fails. |
| Security | Prompt injection is described on a slide. | You run at least eight attacks against your own system and report what happened and what you changed. |
| Who checks the work | An auto-marked quiz, or nobody. | A reviewer steers the capstone at its checkpoints, two markers grade the submission, and two evaluators hold the viva. |
| What the certificate shows | That the course was completed. | That an exam under exam conditions, a double-marked capstone and a recorded viva were all passed. The credential is verifiable. |
Straight answers on time, price, tools and the credential. If yours is not here, ask us and you will have a reply the same working day.
Yes. Bodhih’s LLM Engineering certification, Certified AI/ML Specialist — LLM Engineering (Level 3), is a self-paced online programme for engineers who already write Python and have shipped something. The LLM Engineering course covers retrieval, agents, evaluation, production and governance, and awards the credential only after a certification exam, a double-marked capstone and a recorded viva with two evaluators.
The LLM Engineering course takes about 45 hours of work over 70 days, roughly 10 weeks alongside a full-time job. It is self-paced across 15 stages. The five modules take three to five hours each, the capstone runs for six weeks with five checkpoints, the viva is 30 minutes, and the certification exam is 90 minutes. Capstone marking and viva scheduling follow the submission, so allow some time after it.
The LLM Engineering certification costs ₹64,999 plus GST for an individual, and ₹52,000 per seat for companies. GST at 18% applies to Indian billing addresses and a GST invoice is issued. Individuals pay online by UPI, card or netbanking and get access the moment payment is confirmed. The price covers the diagnostic, five modules, capstone review and marking, the viva, the mock, the certification exam and the credential.
The LLM Engineering course assumes the Certified Applied AI/ML Practitioner (Level 2) credential or equivalent experience. In practice that means you read and write Python confidently and have shipped something that other people use. The course does not teach Python, basic machine learning or what an LLM is. If those are new, Certified AI/ML Foundations (Level 1) and Level 2 are the right places to begin.
The LLM Engineering course teaches every principle tool-agnostically, so the work applies whether your organisation builds on Claude, ChatGPT, Microsoft Copilot, Gemini or another model. Module exercises are written in Python and extend one reference system, with runnable cells in the browser. In the capstone you choose your own embedding model and stack, and you are expected to give a reason for each choice.
The LLM Engineering certification is assessed in three parts, and all three must be passed. The certification exam has 60 scenario-led questions in 90 minutes under exam conditions, with a pass at 70% overall and at least 60% in every domain. The capstone is double-marked on five criteria and passes at 15 out of 20 with no criterion below 2. The 30-minute recorded viva must be judged to meet the standard by both evaluators.
Yes. The LLM Engineering capstone project can be built over your own corpus or over the supplied 48-document corpus. If you use your own, the first checkpoint asks for the corpus with its licence or sponsor confirmation and your anonymisation approach, so the question of permission is settled before you build. A sponsored engineer can therefore build the capstone on a problem the team already has.
The LLM Engineering certification allows one free retake of the certification exam after seven days, and one resubmission of the capstone within 10 working days. The capstone is graded within 10 working days of submission. If the two capstone markers differ by more than three marks or disagree on pass or fail, a third marker decides. A split between the two viva evaluators also goes to a third evaluator.
Yes. The Certified AI/ML Specialist — LLM Engineering credential is verifiable, so an employer or client can confirm that it is genuine, and it is also issued as an Open Badge 2.0. Assessments run on AssessAll, Bodhih’s sister assessment platform. Ninety days after you finish, you re-sit the diagnostic, which shows whether the capability held once the course was over.
Yes. Companies can buy seats on the LLM Engineering course at ₹52,000 per seat for companies, plus GST, which is lower than the individual price. Each engineer builds and defends a capstone individually, so a sponsor receives separate evidence for every seat. Bodhih also delivers the programme as an in-house workshop; there is no public price for that, so talk to us at solutions@bodhih.com or +91 99000 11601.
The best way to learn LLM engineering is to build a real system and have someone qualified challenge your decisions. Look for a course that makes you measure retrieval choices, attack your own guardrails, grow a golden set and model cost at volume, and that has people review the result. Bodhih’s LLM Engineering course does this through five modules, a six-week capstone marked by two people, and a recorded viva.
The LLM Engineering certification is ₹64,999 plus GST for an individual, or ₹52,000 per seat for companies, and access opens the moment your payment is confirmed. The 40-minute diagnostic shows where you stand across design, evaluation, production and governance before module 1. Enrol today and start with the diagnostic. For a team of engineers or an in-house workshop, Bodhih is at solutions@bodhih.com and +91 99000 11601.
Starting records where you stand today. That baseline is what the re-measurement is compared against, and it is never rewritten afterwards.