Shailesh

Founder & Builder. Building deterministic AI systems — agents, self-learning knowledge graphs, applied LLMs, Voice AI. I post what holds up in production.

[Shailesh @spvoiceai]

working on tool scoping, multi-turn failure evaluation, and state consistency in long agent sessions this week.

if that's your world too, say hello.

what are you building?

[Shailesh @spvoiceai]

bigger context windows won't fix agent memory.

stuffing more history into the prompt just moves the forgetting from token 8k to token 200k — the model still can't tell what's load-bearing.

the fix is explicit state, not a wider window.

[Shailesh @spvoiceai]

when one agent in a pipeline fails, do you roll back just its output or every downstream agent that already consumed it?

rolling back everything is safe but expensive. rolling back just the one risks propagating bad state forward.

no clean rule yet.

[Shailesh @spvoiceai]

the timeout needs to live inside the tool call, not the wrapper around it.

a client-side timeout doesn't cancel work in flight. it just retries something that's still running underneath.

[Shailesh @spvoiceai]

@KomalWCode AI finds hard to understand the open source so prefer to write from the scratch

[Shailesh @spvoiceai]

deep in error recovery paths, cost-per-outcome tradeoffs, and memory compaction across long sessions right now.

if any of that's on your desk too, say hello.

what are you building?

[Shailesh @spvoiceai]

a flaky eval usually isn't flakiness.

it's the same prompt producing two valid tool orderings, only one path scored correct.

fix isn't retrying the eval. it's scoring the outcome state, not the exact call sequence.

[Shailesh @spvoiceai]

If you are in #tech let’s #connect

[Shailesh @spvoiceai]

@LiveLoveSEO I told explicitly to use axios but it uses fetch. And not using standard sdk available for the providers.

[Shailesh @spvoiceai]

Hot take:

AI-written code often avoids open-source libraries.

Why?

It’s easier for the model to implement a solution from scratch than to understand the library, its APIs, constraints, and edge cases.

That’s a hidden tax of vibe coding:

More code ≠ better engineering.

[Shailesh @spvoiceai]

@kidtsang Is code written from the scratch by ai better and improve over time? If does this lead to open source libraries dead in future.

[Shailesh @spvoiceai]

@py_poojaa SVN

[Shailesh @spvoiceai]

Does dots run startup?

[Shailesh @spvoiceai]

LAYA open source model

[Shailesh @spvoiceai]

JEV

[Diogo Almeida @CompleteSkeptic]

After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI?

I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev

• 20-200x faster
• 40-400x cheaper (w/ output tokens free)
• Frontier composable intelligence optimized for decisions

AFAICT the shortest path to AI-based economic revolution

[Shailesh @spvoiceai]

most agent framework complexity goes into orchestration. it should go into observability instead.

when an agent fails, wrong tool, right tool bad args, and right call bad timing all look the same from outside. that's the harder problem, not routing between agents.

[Shailesh @spvoiceai]

how many failed attempts before an agent should stop retrying and escalate to a human?

too low and it escalates on transient noise. too high and it burns budget chasing a dead path.

still tuning this per tool, not sure it should be a global number at all.

[Shailesh @spvoiceai]

believed more tools meant a more capable agent.

turned out tool selection accuracy drops once the list gets long — the model second-guesses between similar options.

now: fewer tools, scoped tight to the task, beats a big generic toolbox.

[Shailesh @spvoiceai]

Reasoning Traceability & Audit Logs
You cannot treat your AI's context layer as a black box—especially under strict enterprise governance rules.
The Audit: The database must log reasoning memory. It explicitly tracks which prompt initiated a task, what tools were called, and which specific graph edges were retrieved to build the context window. Complete reproducibility.

[Shailesh @spvoiceai]

Automated Text-to-Graph Generation
Manually structuring knowledge graphs is impossible at scale. Pipelines must automatically turn unstructured silos into structured node-edge networks.
The Pipeline: The GraphDB ecosystem needs native parsers that ingest PDFs, documentation, and chat transcripts—extracting entities, resolving duplicates, and continuously expanding the graph autonomously.

[Shailesh @spvoiceai]

Semantic Layer & Ontology-Driven Inferences
If your data layer doesn't enforce strict domain rules, your AI will fill in the blanks with fiction.
The Guardrails: Building an automated Semantic Layer utilizing standardized ontologies maps explicit data hierarchies and KPIs. AI agents are structurally forced to reason only within the logical boundaries your business defines.

[Shailesh @spvoiceai]

Native Model Context Protocol (MCP) Support
Custom, brittle glue code connecting LLMs to databases is a massive technical debt trap.
The Solution: Native MCP gateway integration allows agent orchestration frameworks (LangChain or LlamaIndex) to securely query the graph database as an ecosystem-vetted tool or "context provider" out of the box.

[Shailesh @spvoiceai]

Hybrid Search: Graph Traversal + Vector Native IndexingBasic keyword or isolated flat chunk retrieval causes LLMs to miss the bigger picture.
True GraphRAG demands a system that natively fuses vector similarity with deterministic graph structures.
The Flow: The AI uses vector math to pinpoint semantic entry nodes, then immediately executes a multi-hop traversal to retrieve a deeply connected context package.