Tesco Bengaluru

Burning Tokens ≈ Burning Money 🔥 If you’re building with LLMs, 𝘁𝗼𝗸𝗲𝗻 𝗲𝗳𝗳𝗶𝗰𝗶𝗲𝗻𝗰𝘆 is no longer an optimization—it’s a requirement. Been researching on how to save tokens from burning. Built this summary I want to share with you all. Save this for future reference. Here are 𝗽𝗿𝗮𝗰𝘁𝗶𝗰𝗮𝗹 𝘄𝗮𝘆𝘀 we’ve been reducing token usage in real systems 👇 1️⃣ Prompt-level optimizations • Shorter prompts, same intent • Constrain outputs: “5 bullets”, JSON only • Use stop sequences to prevent trailing fluff 2️⃣ Context reduction • Send last N messages, not full history • Periodically summarize conversations • In RAG: retrieve only relevant snippets, not entire docs 3️⃣ Retrieval & RAG tuning • Smaller chunk sizes • Lower top_k • Hybrid search + re‑ranking → pass only top results to LLM 4️⃣ Output token control • Ask for answers only, no explanations • Enforce schemas (JSON, enums, yes/no) • Don’t use LLMs for deterministic tasks (sorting, formatting, filtering) 5️⃣ Model choice & routing • Start with smaller/cheaper models • Escalate to larger models only when confidence is low 6️⃣ Caching (huge wins) • Prompt + response caching (FAQs, policies) • Semantic caching for “similar” prompts 7️⃣ Instruction refactoring • Keep system prompts minimal and reusable • Move validations & rules to code • Use tokens only where intelligence is required 8️⃣ Workflow decomposition • Break giant prompts into smaller chained steps • Anything deterministic → code > LLM 9️⃣ Training / fine-tuning • Encode behavior into weights, not prompts • Reuse prompt skeletons for predictable token usage 🔟 Monitoring & feedback loops • Hard input/output token limits per request type • Track which features burn tokens the most LLMs are powerful. They’re also expensive. Treat tokens like a first‑class resource. Use LLMs for inferencing not deterministic things which coding or CPU can do. Curious—what’s been your biggest token burner so far? #AI #LLMs #GenerativeAI #TokenOptimization #AIEngineering #RAG #CostOptimization #SystemDesign #AIatScale

𝗥𝗲𝘁𝗮𝗶𝗹 𝗜𝗻𝘀𝗶𝗴𝗵𝘁 Retail is not about selling products. It’s about influencing behaviour. Pricing. Placement. Promotions. Loyalty programs. Everything is designed to shape decisions. The best retailers don’t just sell— They guide choices. They remove friction. They create habits. Once you understand this— You stop “selling” And start engineering demand. #Retail #RetailStrategy #ConsumerBehavior #MarketingStrategy #LoyaltyPrograms #CustomerExperience

Proud of the results on this one. Strong alignment across the collective team and their vision to get this space lease-ready. Industrial repositioning continues to be a core focus for us.

𝗗𝗶𝘀𝘁𝗶𝗹𝗹𝗲𝗱 𝗞𝗻𝗼𝘄𝗹𝗲𝗱𝗴𝗲 #20 I recently explored this LLM plugin called "𝗖𝗮𝘃𝗲𝗺𝗮𝗻" for AI agents by Julius Brussee . The main goal of Caveman is to reduce tokens. But why? and how? You see when you converse with AI, it takes the context of entire conversation 𝙣𝙤𝙩 𝙟𝙪𝙨𝙩 𝙤𝙣𝙚 𝙢𝙚𝙨𝙨𝙖𝙜𝙚. It's like [prompt1 + response1 + prompt2 + response2 + ... + current prompt] => LLM => response So obviously to reduce money/token burn, and be more efficient we have to reduce the tokens or conversation length. But how can we do it? LLMs can understand the context with the key words in messages, we don't need fillers like is, was, this, that, and etc. When the text chunks are stored in vectors, the major weightage comes from the key words. So as long as they are there, it could gather the understanding right., and in fact it's better, which a research reported. Because the focus is on the core semantic piece of text and the chances of wandering is less. To do this, it is quite interesting, it 𝘁𝗮𝗸𝗲𝘀 𝘁𝗵𝗲 𝗵𝗲𝗹𝗽 𝗼𝗳 𝗟𝗟𝗠 - who is the best to understand some text, and compress it without meaning loss. This requires some tokens, but on a holistic level, this approach saves lots of tokens. It sits as a middleware between User and LLM. User → prompt → compressed prompt → LLM → compressed response → User So when you see the conversation history, you will see compressed ones only., which in turn gets sent to LLM. The benefit keeps compounding. Less text to fewer tokens, not just tokens, speed of delivery as well! Caveman repo link - https://lnkd.in/g3s34eV8 Follow for more interesting ones! Hope this helped. #AI #LLM #Engineering #Productivity #Agents

𝗦𝗘𝗢 𝗶𝘀 𝗻𝗼𝘁 𝗯𝗿𝗼𝗸𝗲𝗻. 𝗬𝗼𝘂𝗿 𝘀𝘁𝗿𝗮𝘁𝗲𝗴𝘆 𝗶𝘀. Most companies: Chase keywords Ignore intent Measure traffic — not revenue Search is not about ranking. It’s about relevance + conversion. The brands that win are not the ones with the most traffic… They are the ones that turn search into business outcomes. #SEO #DigitalMarketing #SearchMarketing #GrowthStrategy

Everyone wants to be “AI-first.” Almost no one wants to fix the data. That image says more than any strategy deck ever could. Top panel: AI-first. A leaking boat. People panicking. Water rushing in faster than it can be removed. Bottom panel: Data-first. Same boat. Same team. But now they’re rowing in sync, actually moving forward. Same ambition. Completely different foundation. ⸻ This is what most enterprise roadmaps actually look like: → Agents running on pipelines nobody owns or maintains → Dashboards pulling from multiple systems, each telling a different story → A “data lake” that quietly turned into a data swamp → Automation that doesn’t improve outcomes — it just scales bad ones faster The issue isn’t AI. It never was. ⸻ Before asking “How do we use AI?”, ask: Do we have duplicate records? Do we know where our data comes from (lineage)? Is there clear ownership? Is anyone accountable for fixing issues? Do we have any real governance? If the answer is no to most of these, AI won’t save you. ⸻ Clean data isn’t a technical problem. It’s a prioritization problem disguised as an IT task. It’s slow. It’s unglamorous. And that’s exactly why it gets ignored. ⸻ But someone has to do that work. Fix the foundation first. Until then, you’re just adding more people to a sinking boat… and calling it progress.

Excited to start a new chapter at Accenture as AI Decision Science Manager. Inspired by its culture of reinvention and focus on driving AI-powered decision-making. Looking forward to building meaningful impact. #Accenture #AI #Reinvention

Still buzzing off the energy from Tesco Technology TPM offsite! Really enjoyed connecting with TPMs across different domains, fresh perspectives, shared challenges, and lots of learnings from each other. I also volunteered to present on resourcing & budgeting along with Raphael R . Huge thanks to Preeja Kuriakose and Deepa Senan for the support. We pulled the presentation together in a short time, truly reflecting our Win Together spirit. Excited to take this forward. #TPM #TechnicalProgramManager #ProgramManagement #TechCareers #WomenInTech #Leadership #ProjectManagement #CareerGrowth #Tescotechnology

Stop Reporting. Start Driving Decisions The difference between a Manager and a Leader is very simple: Managers report numbers. Leaders change outcomes. If your work stops at reporting, trust me you're replaceable. If your work drives decisions, you are valuable. In today's world, data is everywhere. Insight is rare. Action is what creates impact. #Leadership #Management #DataDriven #DecisionMaking #BusinessLeadership #LeadershipDevelopment #Strategy