@agupta

AI/ML Eng | Data Engineering | Generative AI | Cloud & Databricks | Building intelligent systems that scale

[@agupta @anuragtech5]

@7Anoop7 @_vmlops more like 15

[@agupta @anuragtech5]

@buildwtim Followed back mate

[@agupta @anuragtech5]

@danialhasan would be interested !

[@agupta @anuragtech5]

@iHarnoorSingh Buy a Duo man

[@agupta @anuragtech5]

@MicrosoftLearn Guide us how ?

[@agupta @anuragtech5]

@Ozacle23 Did you build this ? Are you a developer?

[@agupta @anuragtech5]

@DeRonin_ Why peddle lies man 6 months really ?

[@agupta @anuragtech5]

@oprydai hmm..where shall I begin with...Howard Stark ?

[@agupta @anuragtech5]

@asmah2107 Let’s go 👋

[@agupta @anuragtech5]

@luisnomad Nice to know you @luisnomad let’s connect 👋

[@agupta @anuragtech5]

@oprydai How far away are we from terminator style killing machines

[@agupta @anuragtech5]

@AkshayNarisetti wow ! congratulations

[@agupta @anuragtech5]

Built an MCP server for FinSight. Claude Desktop now calls my SEC filing RAG pipeline as a tool — discovers it, queries it, gets grounded answers.

Hardest part wasn't the code. It was a print() statement corrupting the JSON-RPC protocol and a 30s model load causing startup timeouts. Real integration problems, not tutorial problems.

Two tools: one returns answers, one returns raw chunks. Different agents want different things.
https://t.co/fm7LuWQJSd

#AIEngineering #MCP #RAG

[@agupta @anuragtech5]

Added RAGAS evaluation to my RAG pipeline. Immediately found the problem: dense financial tables embed as meaningless numbers — retriever can't find what it can't understand.

Fix: contextual chunking. LLM labels each chunk at ingestion. But labels in the generator context killed faithfulness. Had to separate embedding representation from generation input.

Context recall: 0.76 → 0.86. Lesson: if you're not measuring retrieval and generation separately, you're guessing.

https://t.co/fm7LuWQJSd

#AIEngineering #RAG #LLMOps

[@agupta @anuragtech5]

@dharmeshba @GrowthX_Club ha ha this is awesome..the blank expression without any emotion just reading the news..

[@agupta @anuragtech5]

Built a two-stage retrieval pipeline in FinSight this week.

The problem: semantic search returns chunks that are "close" in vector space, but not necessarily the most relevant. Ask "Did Apple's services revenue decline?" and you get chunks containing "services" and "revenue" — but they might be about revenue growth, not decline. The embedding matched keywords, not meaning.

The fix: retrieve wide, then rerank precisely.

Stage 1 — Hybrid search (BM25 + vector + Reciprocal Rank Fusion) casts a wide net. Keyword matching catches exact financial terms that embeddings miss. Vector search catches semantic meaning. RRF merges both without needing score calibration.

Stage 2 — A cross-encoder reranker (BAAI/bge-reranker-base) takes the top 20 candidates and scores each query-chunk pair jointly — full cross-attention, not separate embeddings. Returns the top 5.

The key insight: bi-encoders encode query and document separately — fast but lossy. Cross-encoders read them together — slow but accurate. You can't afford cross-encoders on your whole corpus, so you use both: cheap recall first, expensive precision second.

Same architecture Google, Bing, and every production search system uses. Self-hosted model, no API dependency, no token cost.

One honest finding: on a simple revenue query, pure vector search actually returned more relevant chunks than the reranked pipeline. The reranker can only reorder what it receives — if the hybrid retriever's candidate pool is worse for a particular query type, reranking can't fix it. That's why evaluation matters more than any single technique.

Next up: RAGAS evaluation pipeline to measure impact across 20+ queries with real numbers.

https://t.co/fm7LuWQJSd

#BuildInPublic #AIEngineering #RAG #LLM #InformationRetrieval

Have a look , suggestions are welcome !

[@agupta @anuragtech5]

@HamelHusain @sh_reya Aah this is gold perfect for a weekend learning !

[@agupta @anuragtech5]

Been heads down on FinSight this week 🔍

Production RAG over Apple SEC filings.
Shipped 5 features:

🔴 Redis scan_iter() — non-blocking cache keys
📊 store=True — OpenAI completions logging
🛡️ Rate limiting — 429 after 10 req/min
🔄 Retry logic — exponential backoff
🔍 Hybrid search — BM25 + Vector + RRF

Proof hybrid search works 👇

Query: "AAPL stock repurchase program"
→ Semantic only: 5 chunks
→ Hybrid: 8 chunks
→ BM25 found 3 pages semantic missed entirely

Not a demo. A production system.

https://t.co/fm7LuWQJSd

#BuildInPublic #AIEngineering #RAG #LangChain #Python

[@agupta @anuragtech5]

@Suryanshti777 Roadmap