Zomato: ML Engineer, Campus (2025, Offer)

from langchain_community.document_loaders import PyPDFLoader
from langchain_text_splitters import RecursiveCharacterTextSplitter
from langchain_community.vectorstores import FAISS
from langchain_openai import OpenAIEmbeddings, ChatOpenAI
from langchain_core.prompts import ChatPromptTemplate
from langchain_core.runnables import RunnablePassthrough
from langchain_core.output_parsers import StrOutputParser

docs = PyPDFLoader("menu_policy.pdf").load()
chunks = RecursiveCharacterTextSplitter(chunk_size=800, chunk_overlap=100).split_documents(docs)
retriever = FAISS.from_documents(chunks, OpenAIEmbeddings()).as_retriever(search_kwargs={"k": 4})

prompt = ChatPromptTemplate.from_template(
    "Answer only from the context. If unsure, say you don't know.\n\nContext:\n{context}\n\nQuestion: {question}")
fmt = lambda ds: "\n\n".join(d.page_content for d in ds)
chain = ({"context": retriever | fmt, "question": RunnablePassthrough()}
         | prompt | ChatOpenAI(model="gpt-4o-mini", temperature=0) | StrOutputParser())
print(chain.invoke("What is the refund policy for late deliveries?"))

• Mention what you would add with more time: metadata and citations, a re-ranker, evaluation (RAGAS) and streaming.

• Word2Vec:
• Static: one vector per word type, learned with a shallow skip-gram or CBOW network.
• No context ("bank" always has the same vector), out-of-vocabulary words are a problem, and it is cheap.
• BERT:
• Contextual: the vector depends on the whole sentence through deep bidirectional self-attention.
• Uses sub-word (WordPiece) tokens, so there is no OOV problem.
• Captures syntax and senses, but costs more to compute.
• For sentence-level similarity, plain BERT [CLS] vectors are weak; use sentence-transformers fine-tuned contrastively.

• Problems:
• Words broken into many pieces ("un", "##believ", "##ably") dilute meaning across tokens and use up the context length.
• Rare names, Hinglish ("bahut tasty") and dish names ("paneer-tikka") fragment badly.
• NER: entity boundaries must be re-aligned from sub-words to words; labelling only the first sub-word, or aggregating, adds noise.
• Sentiment: negations and intensifiers split oddly; more tokens means more sequence and a higher chance of truncation.
• Fixes:
• A domain or multilingual tokenizer (or extend the vocabulary with domain tokens and fine-tune).
• Normalise text.
• Use word-level aggregation (first or mean sub-word) for NER.
• Choose models trained on Indic and code-mixed data.

• Sinusoidal (fixed): no parameters; it encodes relative offsets through sin/cos phase and can extrapolate somewhat to longer sequences.
• Learned absolute (ViT patch positions, BERT): flexible and data-driven, but limited to the trained length or resolution. Changing image size requires interpolating the position grid.
• Convolutional or patch stems bring implicit positional information (local inductive bias, zero-padding effects), which some models use instead of, or as well as, explicit embeddings, as in ConvNeXt-style or CvT designs.
• Modern LLMs use RoPE or ALiBi (relative) for better length generalisation.

• Speech-to-text (Whisper, Conformer):
• A smaller or distilled model (distil-whisper).
• Streaming / chunked decoding with VAD to skip silence.
• CTranslate2 or faster-whisper, batching, GPU FP16 or INT8.
• Text-to-speech (FastSpeech, VITS):
• Non-autoregressive models for speed.
• A streaming vocoder.
• Cache frequent phrases.
• Quantization:
• INT8 usually costs under 1% WER or MOS.
• 4-bit can hurt accents, noisy audio and prosody noticeably.
• Measure WER (speech-to-text) and MOS or MCD (text-to-speech) on your data (Hindi or Hinglish, noisy restaurant audio) before and after.
• Use calibration data from the real domain.

1. Measure first: trace every step (LangSmith or Langfuse) and find where it fails. Common culprits are wrong tool choice, loops and slow steps.
2. Routing: a cheap classifier or small LLM sends simple intents (order status) to deterministic flows. Only complex cases reach the full agent.
3. Tools: tighter descriptions and schemas, fewer and more specific tools, validation of arguments, idempotent APIs.
4. Memory and context: summarise long chats, retrieve only the relevant policy chunks, and cache user and order lookups.
5. Reliability: a maximum number of steps, timeouts, retries with backoff, fallback models, and human handoff for refunds above a threshold.
6. Evaluation: a trajectory test set; measure resolution rate, average steps, latency, cost per conversation and CSAT.

• Be honest and positive: "Yes, I'm flexible on location (Gurugram or Bengaluru). I'm most excited about applied ML that reaches users fast, like search, recommendations and support automation at Zomato's scale."
• Show that you know the product: delivery-time prediction, restaurant ranking, and the AI support assistant.