EdYoda, ITPL Road, Bangalore (2026)

23/09/2026

👇Most LangChain tutorials teach you the wrong 80 percent. 🧱

🔥Seven concepts carry almost all the weight once something is actually running in production. The rest you can look up when you need it.

1️⃣ LCEL ⛓️
prompt | model | parser, composed into one chain. Streaming, batching and async come with it instead of being written by hand each time.

2️⃣ Tools and function calling 🔧
The model decides which function to call. Your system executes it. Credentials never leave your side, which is the whole reason this pattern exists.

3️⃣ RAG 📚
The answer is in your HR policy, not in the model. Retrieve the relevant chunks at runtime and hand them over before generation.

4️⃣ StateGraph 🕸️
Nodes, edges, state. This is where a chain becomes a workflow. Low risk goes straight to an offer letter, high risk routes to a human underwriter, and both paths carry the same state.

5️⃣ Checkpointers 💾
A loan application pauses for compliance approval overnight and resumes tomorrow on the same thread_id. Persist the graph state and you get pause, resume, replay and time travel. Without it, an interruption means starting over.

6️⃣ Gateway 🚦
Three providers, one control layer. Failover, rate limits, caching, budgets, routing. Tools like LiteLLM also let you set spend limits per department, which is usually finance's first question.

7️⃣ Observability 👀
Your agent denied a claim. You need the retrieved chunks, the tool arguments and the final decision in one trace. If you cannot trace it, you cannot debug it.

⚡Worth knowing: since LangChain 1.0, create_agent is the recommended entry point and compiles down to a LangGraph graph. LCEL is still the right tool for linear pipelines.

And the honest version: 6 and 7 are premature on day one. One provider and a print statement is fine until it is not.

⭐ Concepts 1 to 3 get a demo working. 4 to 7 are what keep it running at 2 am.

Save this 🔖 and send it to whoever is about to start building.

Which of the seven did you learn the hard way? 👇

Follow EDYODA for AI production engineering, not tutorials.

22/09/2026

👉Your LLM is overkill for most decisions your product makes.

🔥Meet Jev from TypeSafe AI. It's not a chatbot. It doesn't write text. It returns a decision and a probability, in one shot.

1️⃣ The gap: classifiers are fast but narrow. LLMs generalize but are slow and costly at scale. Jev claims both: general AND fast.
2️⃣ The input: instructions (the question), criteria (allowed answers), state (the context). JSON in, JSON out.
3️⃣ Why it's fast: it's non-autoregressive. No token-by-token generation. Claimed 20 to 200x faster and 40 to 400x cheaper than comparable LLMs.
4️⃣ Three decision types: Noul (yes/no), Choice (up to 255 options), Score (2 to 10 levels).
5️⃣ Not for: chat, writing, or multi-step reasoning. It can still misclassify. Use the probability to route low-confidence calls to a human.

⭐ The shift isn't smarter AI. It's good-enough AI at classifier speed and price.

Save this for your next build. Share it with the engineer who pipes every if/else through GPT.

Follow EDYODA for AI, explained properly.

Credit: Matt Canham's "Jev Explained for Normies"

17/09/2026

👇Calling an LLM is not the same as building an application. 🫯

⁉️The gap is four specific problems. This is what LangChain is actually for 👇

1️⃣ The model cannot touch your systems 🔧
"Where is ticket INC12345?" The LLM has no ServiceNow. Your app does. You expose get_ticket_status(ticket_id) as a tool, the model decides when to call it, your code executes it. The model never gets credentials.

2️⃣ Real work is multi-step 🪜
Onboard a vendor means check SAP, create the code, send the NDA, raise a Jira, email finance. Straight line with no branching, LCEL is enough. The moment there is an approval, a retry or a loop, you want a graph.

3️⃣ Your company knowledge is not in the model 📚
The reimbursement policy lives in a PDF, Confluence, SharePoint. Chunk, embed, store, retrieve, then answer. LangChain is the wiring here. The vector DB is the thing actually holding your data.

4️⃣ Decisions need state and clean data 🧠
Leave balance is 5 days, policy says above 3 needs approval, the decision needs both facts alive at the same time. And ServiceNow wants {"status": "resolved"}, not a sentence that happens to contain it. Structured output is not a nice-to-have, it is the contract between your model and your API.

💯One update the diagram does not show. Since LangChain 1.0 last October, create_agent is the front door and it compiles down to a LangGraph graph anyway. AgentExecutor is in maintenance mode until December 2026. If you are copying a 2024 tutorial, you are learning a deprecated API.

🎖️Honest version: if your use case is summarise this text, call the API directly. The abstraction earns its cost at tools, state and multi-provider, nowhere else.

⭐ The model decides. Your system executes. Everything in this carousel lives in that gap.

👍Save this 🔖 and send it to whoever on your team says LangChain is just a wrapper.

🛑Which of the four broke your first agent? 👇

Follow for how this holds up in production.

16/09/2026

👇Delete the customer. The vector is still there. 🔓

⁉️Five checks I would run before any RAG goes near real user data. These are the leaks nobody files as a bug, because nothing errors out.

1️⃣ Redact before you embed 🧼
Once PII is in the vector, it is in the vector. Scan and replace at ingestion, embed only clean text. Presidio is open source and built for this. Redacting only the final answer fixes nothing.

2️⃣ Enforce access at the vector layer 🔐
Prompt instructions are not security. They can be injected. Tag every chunk with customer_id at ingestion and filter at query time. Most teams lock the source DB and leave the vector DB open.

3️⃣ Isolate and encrypt the store 🗄️
"Embeddings are just numbers" is the most expensive sentence in RAG. Vec2Text showed an embedding can be inverted back into the original short text with high accuracy, clinical notes included. Separate collections by sensitivity, encrypt at rest, strip PII from filenames.

4️⃣ Last mile output firewall 🛡️
Clean retrieval can still quote an account number verbatim. Scan the final answer, block or mask, summarise instead of quoting.

5️⃣ Make delete mean delete 🗑️
delete() in HNSW usually just sets a flag. The node stays in the graph until compaction runs. Map user to doc to vector to cache to logs, delete all of it, force the rebuild, keep the audit log.

⭐ Soft delete is not erasure. Your erasure deadline is really your compaction schedule, whether you planned it that way or not.

Save this 🔖 for your next security review and send it to whoever signed your DPA.

🫯Which of these five would fail in your stack today? 👇

Follow for how this stuff behaves in production.

14/09/2026

👇Your RAG demo works. Production is where it quietly lies. 😅

🛑Six months running a RAG agent on WhatsApp across 250+ programs taught me one thing: the model was never the hard part. The system around it was.

✌️5 patterns I wish I had on day one 👇

1️⃣ Multi-Tenant RAG 🔐
Tag tenant ID at ingestion, then filter at all 5 stages. Most teams filter only the vector DB. Rerank is where the leak shows up.

2️⃣ Evaluation Harness ✅
RAG does not crash when it is wrong. It answers confidently. 50 to 100 golden Q&A pairs, run on every deploy. No eval, no prod.

3️⃣ Layered Caching ⚡
"What is your refund policy" and "can I get my money back" are the same question. Cache at edge, pre vector, pre LLM. Cheapest win on the list.

4️⃣ Blue-Green Index 🔄
Build the new index in parallel, validate, atomic cutover. Keep the old one warm for 24h. Do not delete it the same evening.

5️⃣ Data Flywheel 📈
Thumbs down, corrections, low confidence logs. Most teams let all of it evaporate. Filter before you learn from it though, a lot of it is user error.

⭐ Your RAG does not fail loudly. It fails politely, and you find out from a customer screenshot.

🔥Save this for your next architecture review 🔖 and send it to whoever owns your retrieval stack⚡

⁉️Which one bit you first in production? I will go deeper on whichever comes up most 👇

Follow for how this stuff actually behaves in prod.

09/09/2026

👉Six steps to Forward Deployed Engineer🚀

🔥Step 6 is the one that separates FDEs from strong engineers. 🧭

1️⃣ LLM applications & AI agents. LLMs, prompt engineering, LangGraph, tool calling, evals. Build agents you actually shipped, this becomes the repo hiring managers open first, before your résumé.
2️⃣ Multi-model LLMs & enterprise data. RAG, hosted vs open-weight (Llama, Mistral), multi-source data. Why it matters: real customers restrict which vendors you're allowed to use, and their data is never clean. Build for any model, not your favourite one.
3️⃣ AWS, Azure & secure deployment. Secrets management, air-gapped deployment. Pick one cloud and you lose half the field. Fail the security review and it never ships, no matter how good the model is.
4️⃣ Docker, Kubernetes & handoff docs. Containerize with Docker, K8s, Terraform. Then write the runbook. Automation makes it work; docs make it survive without you in the room. FDEs leave.
5️⃣ LLMOps & cost modeling. Tracing, drift detection, cost per 1,000 requests. Everyone shows accuracy in a demo. FDEs defend cost, and cost is why most pilots quietly die at renewal.
6️⃣ AI product management & live calls. Turning vague asks into scope, live: discovery calls, product briefs, SOWs.

⭐ Steps 1–5 are HOW. Step 6 is WHY and HOW MUCH. Most engineers build all five and still can't sit across from a customer and scope the thing.

📌 Save this, it's a build order, not a reading list.

💬 Which step are you weakest on? Drop the number 👇

🔔 Follow EDYODA

07/09/2026

👉An Application Engineer. A Senior Platform Engineer with 18 years at Amex. A Data & AI product leader. Same program. 🤖

✅They didn't switch careers. They pulled agentic AI into the work they already do 👇

🧠 Nanda Kishore, Application Engineer. Covered neural networks and deep learning fundamentals, transformers, CNNs and RNNs, LLMs, diffusion models, RAG, then agentic workflows with LangChain, LangGraph, LangSmith and MCP. Now integrating it with backend systems, APIs and event-driven architecture.

⚙️ Shailendra Boro - Senior Platform Engineer, 18+ years in enterprise IT at American Express. His framing is the sharpest one here: automation → autonomy. Scripts → intelligent agents. Reactive ops → self-healing platforms. He's applying it to internal developer platforms, CI/CD, infra provisioning, and observability at scale.

📊 Ravindra Varma Datla - Data Product & AI Transformation. Using it to embed agentic AI into digital transformation strategy rather than to write agents himself.

⭐ Notice the pattern: none of them needed an AI background going in. What they brought was deep domain context - platform engineering, backend systems, data products - and agentic AI became a multiplier on that, not a replacement for it.

That's the actual advantage experienced engineers have here. You already know where the hard problems live.

📌 Save this if you're weighing whether your experience transfers. It does.

💬 What would you point an agent at first in your current role? 👇

🔔 Follow EDYODA

04/09/2026

👉Everyone says you need a vector database for RAG. Most teams don't. 🐘

⁉️Four checks before you add infrastructure you may never need 👇

1️⃣ Check your real scale. Most RAG apps aren't 100M vectors — they're thousands to low millions of chunks. Under 1M, pgvector gives >95% recall at single-digit milliseconds. Same as Pinecone. From 1M–10M it's still competitive, just needs tuning. If you're under 10M in the next 12 months, you're over-engineering.

2️⃣ Check your cost. At 1M–10M vectors: pgvector ~$30/mo, Pinecone ~$180/mo. A real case at 8M vectors and 500K queries/day: ~$200/mo vs $2,000+/mo. That gap is where four-figure savings come from.

3️⃣ Check your query. Do you need "similar chunks, but only for this tenant"? In pgvector that's one WHERE clause. In Pinecone or Qdrant, recall silently drops when your filter narrows to ~2% of data — HNSW searches first and filters after. If permission filtering matters, Postgres wins.

4️⃣ Know where it breaks. You genuinely do need a dedicated DB past 50M vectors (10M needs 60–100GB RAM; 100M is 600GB+ raw storage), if deletes are frequent (pgvector holds tombstones until a full rebuild), or if you need multi-region latency and zero HNSW tuning. pgvectorscale extends Postgres to ~50M on disk if you'd rather stay.

⭐ For most teams right now, your Postgres is the vector DB. Don't add infrastructure because a tutorial told you to.

📌 Save this — it's the checklist before you pick your RAG stack.

💬 What are you actually running, and at what scale? 👇

🔔 Follow for more.

01/09/2026

👉Most AI engineer roadmaps are a reading list. This one is a build order. 🧭

🚀7 layers, in the sequence they actually matter 👇

1️⃣ LLM application. Tokens, context window, temperature. Prompt engineering. Embeddings and vector DBs. Structured output and function calling. API integration — OpenAI, Claude, Gemini.

2️⃣ Retrieval (RAG). Chunking, hybrid search, reranking. Recall@k and Precision@k. Index freshness — incremental updates and deletions, the part everyone skips. And knowing when fine-tuning beats RAG.

3️⃣ Agents. Tool calling and schema design. Loop prevention and max iterations. Human-in-the-loop. Memory across turns. Orchestration as state machines, not prompt chains.

4️⃣ Evaluation & quality. Golden test sets. LLM-as-judge and its failure modes. Faithfulness checks. Prompt versioning and regression testing. This is the layer that separates a demo from a product.

5️⃣ Production. Model routing and cascading for cost. Prompt caching for latency. Observability and tracing. Multi-provider failover. CI/CD for prompts.

6️⃣ Safety & compliance. Guardrails, PII detection, access control on retrieval, prompt injection awareness. Non-negotiable the moment real users touch it.

7️⃣ Model selection & platform. Prompt vs RAG vs fine-tune vs skip the LLM entirely. Open vs closed models. One cloud platform. Multimodal and MCP.

⭐ Notice what's at layer 4, not layer 7: evaluation.

⚡ Most people learn agents before they can measure whether anything works — then wonder why production breaks silently.

📌 Save this — it's a checklist you'll come back to, not a one-time read.

💬 Which layer are you stuck on right now? Drop the number 👇

🔔 Follow EDYODA

31/08/2026

👉Vector DBs are not "AI databases." That phrase is quietly wrecking how people architect systems. 🗄️

🛑60 seconds, no jargon 👇

✅ What SQL is great at: exact, business-critical questions. tenant_id = 'acme_corp' AND status = 'overdue'. JOINs, permissions, audit trails, transactions. Why every bank and SaaS still runs on it.

❌ Where SQL gets stuck: meaning. One ticket says "private subnet fails," another says "isolated VPC." Same root cause. LIKE '%EKS%' misses the second entirely - SQL compares words, not intent.

🔢 The missing piece: embeddings. Models turn text into 256–4096 numbers. Close numbers = close meanings. Two tickets sharing zero words land right next to each other.

🎯 What a vector DB does - one job: "given this meaning, find the 10 closest meanings." That's it. A search engine for meaning, not a replacement for SQL.

⚠️ Why "AI database" misleads: most vector DBs lack full ACID, native JOINs, and fine-grained permissions. Real apps use both- SQL holds tenants, orders, audit logs. Vector DB finds what's similar.

🔀 Hybrid search is how it actually works: SQL filters first (tenant, type, status). Vector search ranks what's left by meaning. Filter, then rank

💡 Need a separate vector DB? Probably not at first. Already on Postgres? pgvector is enough for most projects under 10M docs. Qdrant or Pinecone past that.

⭐ SQL was never designed to compare meanings. Vector DBs were designed exactly for that. Not a replacement - a teammate.

📌 Save this - the last slide is the whole architecture decision in one image.

💬 pgvector or a dedicated vector DB? Which, and why? 👇

🔔 Follow EDYODA

EdYoda

23/09/2026

22/09/2026

17/09/2026

16/09/2026

14/09/2026

09/09/2026

07/09/2026

04/09/2026

01/09/2026

31/08/2026

Address

Telephone

Website

Alerts

Contact The Business

Shortcuts

Share

Category

Listing takedown request