TC Antony

AI Engineer & Founder at Ragleap Intelligence @ragleap

[TC Antony @antonychacko77]

Nvidia is lifting US tech, but the global picture is mixed.

India closed lower, HDFC Bank dragged the Nifty, and Europe slipped.

US futures point higher.

AI optimism ≠ broad risk appetite.

Track the bigger picture:
https://t.co/6fgry4QrcB #Investing #stockmarket

[TC Antony @antonychacko77]

Nvidia is lifting Asian tech. India is different.

KOSPI +1.5% after Nvidia’s results. India remains cautious as HDFC Bank weighs on sentiment.

AI optimism isn’t moving every region the same way.

Track it: https://t.co/0O1BjsgYqa

#Stocks #StockMarket #Trading #Investing

[TC Antony @antonychacko77]

One design choice in RAG systems that I really like: separating retrieval from generation.

Not every query needs an LLM to generate a final answer.

[RagLeap @RagLeap]

Not every use case needs a generated answer — sometimes you just want
the matching chunks.

ragleap-rag now has a standalone retrieve() method: hybrid dense+sparse

[TC Antony @antonychacko77]

@RagLeap This is an important separation in RAG architecture.
Retrieval and generation don't always need to be coupled. Being able to inspect the retrieved chunks independently makes it much easier to evaluate retrieval quality, debug failures, and build custom generation or agent

[RagLeap @RagLeap]

Benchmarks are only useful if they're real. ragleap-rag's benchmarks
cover ingestion, latency, and cost — run for real, with gaps honestly
labeled rather than backfilled.

We've also documented (and rejected) hype pitches with fabricated

[TC Antony @antonychacko77]

One thing I’ve learned from building AI infrastructure:

A feature claim is not the same thing as a verified capability.

“Supports 28 formats.”

“Works with 6 vector backends.”

“Supports 12+ providers.”

[RagLeap @RagLeap]

Every number on the README is backed by a real, reproducible check
in CI — 28 document formats, 6 vector backends, 12+ named providers,
100% opt-in features (reranking, memory, cache — nothing forced).

Not marketing copy. Verified, or it doesn't go in the docs.

[TC Antony @antonychacko77]

@RagLeap This is a principle more open-source projects should follow.

A README can claim almost anything, but reproducible CI checks turn those claims into something contributors can actually verify.

Especially for RAG libraries, where supported formats, providers, backends,

[TC Antony @antonychacko77]

nowledge-graph-augmented retrieval is one of the areas in RAG that I think deserves more attention.

A vector database answers:

→ “What information is semantically similar to this query?”

A knowledge graph can answer:

→ “How are these entities connected?”

[RagLeap @RagLeap]

Roadmap update: ragleap-graph shipped.

It's an optional package that composes with ragleap-rag to add
knowledge-graph-augmented retrieval — structured entity relationships
alongside vector search, not instead of it.

[TC Antony @antonychacko77]

@RagLeap Knowledge graphs are most useful when they complement retrieval rather than trying to replace it.

Vector search is great at semantic similarity, while a graph can provide explicit relationships between entities.

[TC Antony @antonychacko77]

Low latency is an architectural decision, not a last-minute optimization.

In RAG systems, latency can accumulate quickly:

→ Retrieval
→ Reranking
→ Database calls
→ LLM generation
→ Tool calls
Optimizing only one component rarely solves the whole problem.

[RagLeap @RagLeap]

Speed isn't bolted on after the fact — it's structural:

- Async-first — every method has a non-blocking twin
- Streaming — first-token latency, not full-answer wait
- Connection pooling — no per-request Postgres handshake
- No cold-start GPU tax — ONNX CPU reranking,

[TC Antony @antonychacko77]

Security-first is especially important for open-source AI infrastructure.

One thing I’ve learned while working on RAG systems is that security isn’t only about protecting the application layer.

The dependency and software supply chain matter too.

[RagLeap @RagLeap]

Security work most RAG libraries skip:

- SECURITY.md

with a real vulnerability disclosure process
- Auto-regenerated CycloneDX SBOM on every dependency change
- CodeQL + Dependabot active on every PR, not opt-in

Shipped and merged — not a promise.

pip install ragleap-rag

[TC Antony @antonychacko77]

A good reminder that retrieval quality starts before the retriever.

A user's raw question isn't always the best representation of what they actually need to retrieve.

Query rewriting can help in different ways:

[RagLeap @RagLeap]

A raw user question is often a bad search query. ragleap-rag supports
three optional rewriting strategies before retrieval:

- Contextual follow-up resolution
- HyDE (hypothetical document embeddings)
- Multi-query RAG-Fusion

[TC Antony @antonychacko77]

@RagLeap Query rewriting is one of those improvements that can look small but have a big impact on retrieval quality. I especially like keeping these strategies opt-in — different RAG workloads need different retrieval behavior.

[RagLeap @RagLeap]

A raw user question is often a bad search query. ragleap-rag supports
three optional rewriting strategies before retrieval:

- Contextual follow-up resolution
- HyDE (hypothetical document embeddings)
- Multi-query RAG-Fusion

[TC Antony @antonychacko77]

RAG quality isn't just about embeddings. It's about retrieval strategy.

One thing I've learned while building RAG systems is that relying exclusively on dense retrieval can leave important information behind.

Exact terms matter.

[RagLeap @RagLeap]

Dense-only retrieval misses exact keyword matches. Sparse-only misses
semantic meaning. ragleap-rag fuses both — pgvector (dense) + full-text
(sparse) — via Reciprocal Rank Fusion.

Reranking runs on a 23MB ONNX cross-encoder. No torch, no CUDA, no

[TC Antony @antonychacko77]

I built a RAG engine that runs on CPU.

23MB ONNX reranker > 2GB torch model
WhatsApp + Telegram + Discord built in
No agents. No CUDA hell.

GitHub: https://t.co/bqNQKoEW2C

What did I miss? #RAG #LLM #OpenSource

[TC Antony @antonychacko77]

I just published The Bottleneck Isn’t the AI. It’s the Brief. https://t.co/0YV3ZjaGtJ

[TC Antony @antonychacko77]

🎯 Tactical breakdowns ⭐ Key players to watch

Who will reach the Final? France 🇫🇷, Spain 🇪🇸, England 🏴, or Argentina 🇦🇷? Drop your prediction below! 👇⚽
#WorldCup2026 #FIFA2026 #Football #Soccer #AI #SportsAI #MatchPrediction #France #Spain #England #Argentina #WC2026

[TC Antony @antonychacko77]

🏆 The road to the FIFA World Cup 2026 Final is almost complete! ⚽

🇫🇷 France vs 🇪🇸 Spain

🏴 England vs 🇦🇷 Argentina

Analyze every remaining World Cup fixture with AI-powered insights.
Try it free 👉 https://t.co/QB2G2bAzIg

[TC Antony @antonychacko77]

Four teams remain. Two will reach the Final on July 19 at MetLife Stadium.
Don't rely on guesses—use data.
World Cup AI Match Intelligence provides:

✅ Win / Draw / Loss probabilities 📊 Scoreline predictions with confidence % 📈 Team form & momentum analysis