Publicis Sapient: Senior Associate Data Scientist

• Scope: for example, "a product-FAQ assistant for a retail client". Define 20 test questions first.
• Stack:
• An open model through an OpenAI-compatible endpoint (vLLM, TGI or a provider).
• bge-small embeddings and FAISS/Chroma.
• LangChain LCEL.
• FastAPI + a simple Streamlit/Gradio UI.

from openai import OpenAI          # works with OpenAI-compatible open-model servers
client = OpenAI(base_url="http://localhost:8000/v1", api_key="not-needed")
def answer(q, ctx):
    msg = [{"role": "system", "content": "Answer only from the context; cite [n]."},
           {"role": "user", "content": f"Context:\n{ctx}\n\nQuestion: {q}"}]
    return client.chat.completions.create(model="meta-llama/Llama-3.1-8B-Instruct", messages=msg, temperature=0).choices[0].message.content

• Deliverables: a README (architecture, how to run), an evaluation table (accuracy or faithfulness on the 20 questions), the known limitations, and next steps (re-ranker, guardrails).

• For: data control and on-premises deployment, no per-token vendor lock-in, fine-tuning freedom, predictable cost at scale, and region or regulatory compliance.
• Against: lower quality on hard reasoning tasks, ops burden (GPUs, scaling, updates), and safety tuning is your responsibility.
• Pattern: prototype on both, compare on your evaluation set, and choose by quality, cost and latency. A router can mix them: open models for simple queries, a frontier API for hard ones.

• High bias (underfit): the model is too simple. Add features, use a more flexible model, reduce regularisation.
• High variance (overfit): the model is too complex for the data. Get more data, add regularisation or dropout, simplify, use bagging or early stopping.
• Tune with validation curves (vary complexity) and learning curves (vary data size), using cross-validation.

• LLM (reasoning) + tools (APIs, search, code) + memory (short-term state, long-term store) + planning / control loop (ReAct, plan-and-execute) + guardrails (limits, approvals).
• Multi-agent: a supervisor routes tasks to specialists (researcher, coder, reviewer).
• Evaluate on task success, steps, cost and safety violations.

from collections import Counter

def first_unique(s):
    c = Counter(s)
    return next((ch for ch in s if c[ch] == 1), None)

print(first_unique("swiss"))  # w

• Two passes, O(n) time, and O(alphabet size) space.

• Start with the decision and the impact: "The model flags 8% of orders as high-risk and catches 70% of fraud, saving about ₹40 lakh a month."
• Show drivers with plain-language SHAP summaries ("new device + high amount + night-time").
• Be honest about errors: "1 in 5 flags is a false alarm; here's the review process."
• Use visuals (lift chart, before/after), avoid jargon, and end with a recommendation and next step.